Deep learning flood forecasting method considering random hydrological process
By combining the deep learning model and the Markov model, the loss function of the time domain convolutional neural network model is improved, and a deep learning flood forecast model considering random hydrological processes is constructed, which solves the problem that existing methods are difficult to capture local key events and rapid changes, and improves the accuracy and stability of flood forecasting.
Patent Information
- Application Number
- CN202411818673.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-11
- Publication Date
- 2025-04-18
AI Technical Summary
Existing deep learning methods are difficult to effectively capture local key events and rapid changes in flooding during training, resulting in increased difficulty and uncertainty in flood forecasting.
By combining the global timing dependence of deep learning models and the local randomness of Markov models, the loss function of the time domain convolutional neural network model is improved, and a deep learning flood forecast model (TCN-Markov) that considers random hydrological processes is constructed.
It effectively makes up for the shortcomings of traditional deep learning methods in capturing local key nodes and rapidly changing situations in the flood process, improves the accuracy and stability of flood forecasting, and has stronger robustness and adaptability.
Smart Images
Figure CN119849547A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of flood forecasting, and in particular to a deep learning flood forecasting method taking random hydrological processes into consideration. Background Art
[0002] Flood forecasting is an important task in hydrological science. Accurate flood forecasting can effectively reduce the losses caused by natural disasters. Traditional flood forecasting methods mainly rely on physical models and statistical models. These methods make predictions by considering historical hydrological data and meteorological characteristics, but they usually assume that the hydrological process is deterministic and ignore the randomness and uncertainty of the hydrological process. However, there is local randomness in the flood process. In actual situations, the formation process of floods is affected by multiple complex uncertain factors, and the interaction between these factors makes the occurrence of floods highly random. Due to this randomness, it is difficult to accurately predict the occurrence and development of floods by relying solely on traditional deterministic models.
[0003] In recent years, deep learning methods (such as time series convolutional networks) have demonstrated powerful capabilities in time series forecasting and can automatically learn global dependencies in time series data. In particular, deep learning networks can capture long-term dependencies and temporal regularities in flood forecasting, such as long-term trends and changes between variables such as precipitation, flow, and water level. This enables deep learning to comprehensively consider the entire flood process and make more accurate forecasts when dealing with complex problems such as flood forecasting.
[0004] However, existing deep learning methods still have shortcomings when dealing with local key events (such as changes in key turning points of floods) during training. There are often some key nodes in flood forecasting that change significantly in a short period of time, such as when the water level reaches the warning line or the flow surges. These local changes are closely related to the state of the previous period, which determines the development trend of the flood event. These key nodes usually correspond to the critical state of the system. Especially during the peak period of the flood and in rapidly changing scenarios, this local randomness of the flood and the changes in key nodes are often difficult to be captured by the deep learning model alone, resulting in a sharp increase in the difficulty and uncertainty of the prediction.
[0005] Therefore, how to impose local dependency constraints during deep learning model training has become a key issue in improving flood forecasting accuracy. Summary of the invention
[0006] In order to overcome the shortcomings of the prior art, the present invention provides a deep learning flood forecasting method taking into account random hydrological processes.
[0007] The technical solution of the present invention to solve the above technical problems is:
[0008] A deep learning flood forecasting method considering random hydrological processes includes the following steps:
[0009] Step S1: Collect the hydrological and meteorological data of the basin during the flood period, and standardize the collected hydrological and meteorological data;
[0010] Step S2: Classify the flood process according to the hourly flow rate, and use the Markov model to construct the state transition probability index of different flood process levels;
[0011] Step S3: Using the time domain convolutional neural network model as the basic model, the loss function in the time domain convolutional neural network model is improved by the state transition probability index, so as to construct a deep learning flood forecasting model (TCN-Markov) considering random hydrological processes;
[0012] Step S4: dividing the standardized hydrological and meteorological data into a training set and a validation set, training the constructed deep learning flood forecasting model through the training set and validating the deep learning flood forecasting model through the validation set, to obtain a trained deep learning flood forecasting model;
[0013] Step S5: Forecast floods using the trained deep learning flood forecasting model.
[0014] A preferred embodiment of the present invention is that in step S1, the hydrological and meteorological data of the basin flood period include five-dimensional data of hourly water level, hourly flow, rainfall, evaporation and temperature for many years. The water level and flow data directly reflect the flood process of the basin flood period, the rainfall and evaporation data reflect the indirect impact on the basin flood process from the water vapor circulation intensity, and the temperature data reflects the potential impact on the basin flood process.
[0015] In a preferred embodiment of the present invention, in step S1, the data standardization processing method selects the maximum minimum standardization method, and the specific formula is as follows:
[0016]
[0017] In the formula, x i 、x i ′ are the data before and after standardization, respectively, x min 、x max are the minimum and maximum values of the data series, respectively.
[0018] In a preferred embodiment of the present invention, in step S2, the flood process is classified according to the hourly flow, specifically into base flow period (L1), flood period (L2), flood peak period (L3), water receding period (L4) and recovery period (L5); and a precondition is added to assist in determining the flood process classification at a certain moment; wherein,
[0019] The prerequisite for judging whether a flood process at a certain moment is the base flow period is that the previous period is the base flow period or the flood period;
[0020] The precondition for judging whether a flood process at a certain moment is a flood period is that the previous period is a base flow period or a flood period;
[0021] The prerequisite for judging whether a flood process at a certain moment is a receding period is that the previous period is a flood peak period or a receding period;
[0022] The prerequisite for judging whether a flood process at a certain moment is a recovery period is that the previous period is a flood peak period, a receding period or a recovery period.
[0023] In a preferred embodiment of the present invention, in step S2, a Markov model is used to construct a state transition probability index of different flood process levels, and the specific formula is:
[0024]
[0025] Where: P ij For state L i Transfer to state L j The probability of Q ij For state L i Transfer to state L j The frequency, Q i For state L i The frequency of occurrence; among them,
[0026] The state transition probability index of the base flow period includes: 11 , base flow period turns to flood period P 12 , the rapid transition from base flow to peak flow caused by extreme rainfall 13 ;
[0027] The probability index of state transition during the flood period includes: the transition from the flood period to the base flow period due to insufficient precipitation P 21 , the flood season turns to the flood seasonP 22 , the flood season turns to the peak seasonP 23 ;
[0028] The state transition probability index during the flood peak period includes: the flood peak period turns to the flood peak period P 33 , flood peak period turns to receding periodP 34 , the flood peak period quickly turns to the recovery period due to human intervention in flood discharge P 35 ;
[0029] The state transition probability index of the water-reduction period includes: 44 , the water withdrawal period turns to the recovery period P 45 ;
[0030] The state transition probability index of the recovery period includes: 55 .
[0031] In a preferred embodiment of the present invention, in step S3, the time domain convolutional neural network model includes 3 residual blocks, 1 Flatten layer and 1 Dense fully connected layer, wherein each residual block contains 3 convolutional layers, and the number of neurons in the 3 convolutional layers is 128, 128 and 64 respectively; the number of neurons in the Dense fully connected layer is 64; the activation function is Relu, the optimizer is Adam, the learning rate is 0.001, and the loss rate is 0.2.
[0032] In a preferred embodiment of the present invention, in step S3, the baseline loss function of the time-domain convolutional neural network model is:
[0033]
[0034] Where: Loss p is the benchmark loss function, Y i is the true value; is the simulated value, n is the number of samples;
[0035] The benchmark loss function of the time-domain convolutional neural network is improved by the Markov state transition probability index. The specific improvement formula is:
[0036]
[0037] In the formula, Loss n is the loss function of the Markov constraint, Y i is the true value; is the simulation value, P i is the state transition probability index corresponding to the true value; P i ′ is the state transition probability index corresponding to the simulation value, and n is the number of samples.
[0038] In a preferred embodiment of the present invention, in step S4, 80% of the standardized hydrological and meteorological data are used as a training set, and the remaining 20% are used as a validation set.
[0039] An electronic device comprises a central processing unit and a memory, wherein the central processing unit is used to call and run a computer program stored in the memory to execute the steps of the deep learning flood forecasting method considering random hydrological processes.
[0040] A computer-readable storage medium stores a computer program implemented according to the deep learning flood forecasting method considering random hydrological processes in the form of computer-readable instructions. When the computer program is called and executed by a computer, the steps included in the corresponding method are executed.
[0041] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0042] 1. The deep learning flood forecasting method considering random hydrological processes of the present invention combines the global temporal dependency of the deep learning model and the local randomness of the Markov model, thereby effectively making up for the shortcomings of traditional deep learning methods in capturing local key nodes and rapidly changing situations in the flood process.
[0043] 2. The deep learning flood forecasting method considering random hydrological processes of the present invention optimizes the loss function of the deep learning model by introducing the state transition probability constraints of the random hydrological process (Markov model), so that the constructed deep learning flood forecasting model can better focus on the sudden random changes and key nodes in the flood process, thereby improving the prediction accuracy and stability of the model, and having stronger robustness and adaptability under complex hydrological conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 This is a flow chart of the deep learning flood forecasting method considering random hydrological processes of the present invention.
[0045] Figure 2 This is the location map of the Hankou hydrological station in the middle reaches of the Yangtze River.
[0046] Figure 3 This is the main structural diagram of the deep learning flood forecasting (TCN-Markov) model in the present invention.
[0047] Figure 4 This is the main structural diagram of the CNN forecast model of the Hankou hydrological station.
[0048] Figure 5 This is the main structure diagram of the LSTM forecasting model of the Hankou hydrological station.
[0049] Figure 6 This is the flood flow simulation result of the TCN-Markov model at the Hankou hydrological station.
[0050] Figure 7 This is the flood flow simulation result of the CNN model at the Hankou hydrological station.
[0051] Figure 8 This is the flood flow simulation result diagram of the LSTM model at the Hankou Hydrological Station. DETAILED DESCRIPTION
[0052] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.
[0053] Example 1
[0054] Wuhan, where Hankou Station is located, is an important city in the middle reaches of the Yangtze River. It has a dense economy and population and is one of the core areas of the entire Yangtze River Basin. Flood disasters have a particularly significant impact on the region, especially on agricultural production, transportation, and urban safety. Therefore, this paper takes Hankou Station as a research example (e.g. Figure 2 As shown in the figure, a deep learning flood forecasting model (TCN-Markov) as well as the traditional CNN forecasting model and LSTM forecasting model are constructed to simulate the flood flow at Hankou Station.
[0055] See also Figure 1 The deep learning flood forecasting method considering random hydrological processes of the present invention comprises the following steps:
[0056] Step S1: Collect the hydrological and meteorological data of the basin during the flood period, and standardize the collected hydrological and meteorological data;
[0057] Specifically, the flood excerpt data of the Hankou station in the middle reaches of the Yangtze River from 2008 to 2020, including water level, flow, rainfall, evaporation, and temperature, are collected and sorted, and these flood excerpt data are standardized; the flood process of the Hankou station of the Yangtze River is directly reflected through water level and flow data, the rainfall and evaporation data reflect the indirect impact on the flood process of the Hankou station of the Yangtze River from the intensity of water vapor circulation, and the temperature data reflects the potential impact on the flood process of the Hankou station of the Yangtze River;
[0058] The maximum and minimum standardization method is selected as the standardization processing method for flood excerpt data. The specific formula is as follows:
[0059]
[0060] In the formula, x i 、x i ′ are the flood excerpt data before and after standardization, respectively, and x min 、x max are the minimum and maximum values of the data series, respectively.
[0061] Step S2: Classify the flood process according to the hourly flow rate, and use the Markov model to construct the state transition probability index of different flood process levels;
[0062] According to the hourly flow rate of Hankou Station, the flood process is classified into base flow period (L1), flood period (L2), peak period (L3), receding period (L4) and recovery period (L5). Because the flow range of some periods is the same, it is necessary to add preconditions to assist in judging the classification of the flood process at a certain moment. Among them, the precondition for judging that the flood process at a certain moment is the base flow period is that the previous period is the base flow period or flood period; the precondition for judging that the flood process at a certain moment is the flood period is that the previous period is the base flow period or flood period; the precondition for judging that the flood process at a certain moment is the receding period is that the previous period is the peak period or receding period; the precondition for judging that the flood process at a certain moment is the recovery period is that the previous period is the peak period, receding period or recovery period; as shown in Table 1:
[0063] Table 1 Flood process classification table of Hankou station
[0064]
[0065] The Markov model is used to construct the state transition probability index of the flood process level at Hankou Station. The specific formula is:
[0066]
[0067] Where: P ij For state L i Transfer to state L j The probability of Q ij The flood sequence is composed of state L i Transfer to state L j The frequency, Q i The flood sequence is composed of state L i Frequency of occurrence.
[0068] Step S3: Using the time domain convolutional neural network (TCN) model as the basic model, the loss function in the time domain convolutional neural network (TCN) model is improved by the state transition probability index, so as to construct a deep learning flood forecasting (TCN-Markov) model considering random hydrological processes;
[0069] Since the time-domain convolutional neural network (TCN) model is both efficient and stable when processing multi-feature data and time series data, and can parse the global dependencies of data, it is selected as the basic network model. The time-domain convolutional neural network (TCN) model includes 3 residual blocks, 1 Flatten layer and 1 Dense fully connected layer, where each residual block contains 3 convolutional layers; the hyperparameters are set as follows: the number of neurons in the 3 convolutional layers in each residual block is 128, 128 and 64 respectively; the number of neurons in the Dense fully connected layer is 64; the activation function is Relu, the optimizer is Adam, the learning rate is 0.001, and the loss rate is 0.2;
[0070] The time-domain convolutional neural network (TCN) model guides the training model through the back propagation of the loss function. The selected benchmark loss function is the mean square error (MSE), and the specific formula is:
[0071]
[0072] Where: Loss p is the benchmark loss function, Y i is the true value; is the simulated value, n is the number of samples;
[0073] The loss function in the time domain convolutional neural network (TCN) model is improved by using the Markov state transition probability index to enhance the model's local dependence on the flood process of the Hankou station and its ability to capture key nodes, and then a deep learning flood forecasting model (TCN-Markov) considering random hydrological processes is constructed. The specific improvement formula is:
[0074]
[0075] Where: Loss n is the loss function of Markov constraint; Y i is the real value of the flow rate at Hankou Station; is the simulated value of the flow at Hankou station; P i is the state transition probability index corresponding to the true value of the flow at Hankou station; P i ′ is the state transition probability index corresponding to the simulated value of the flow at Hankou station; n is the length of the flood sequence at Hankou station; (1+|P i -P i ′|) 2 The amplification model focuses on the local randomness and key nodes of the flood process at Hankou Station. The specific structure of the TCN-Markov model is as follows: Figure 3 shown.
[0076] Step S4: Divide the standardized hydrological and meteorological data into a training set and a validation set, train the constructed deep learning flood forecasting (TCN-Markov) model through the training set and verify the deep learning flood forecasting (TCN-Markov) model through the validation set to obtain a trained deep learning flood forecasting (TCN-Markov) model; forecast floods through the trained deep learning flood forecasting (TCN-Markov) model; in this embodiment, 80% of the standardized hydrological and meteorological data is used as a training set, and the other 20% is used as a validation set.
[0077] In order to verify the performance of the deep learning flood forecasting (TCN-Markov) model of the present invention, the simulation results of the TCN-Markov model of the present invention and the traditional model are compared and analyzed by constructing the traditional CNN and LSTM flood forecasting models of Hankou Station; specifically:
[0078] (1) Construct the traditional CNN prediction model and LSTM prediction model, where:
[0079] The CNN prediction model includes 3 convolutional layers, 3 maximum pooling layers, 1 Flatten layer and 1 Dense fully connected layer; the number of neurons in the convolutional layer is 256, 128, and 128 respectively; the number of neurons in the Dense fully connected layer is 64; the activation function is Relu, the optimizer is Adam, the loss function is MSE, the learning rate is 0.001, and the dropout rate is 0.2. The main structure of the model is as follows Figure 4 As shown;
[0080] The LSTM prediction model consists of three LSTM layers and one Dense fully connected layer. The number of neurons in the LSTM layer is 256, 128, and 128 respectively; the number of neurons in the Dense fully connected layer is 64, the activation function is Relu, the optimizer is Adam, the loss function is MSE, the learning rate is 0.001, and the loss rate is 0.2. The main structure of the model is as follows: Figure 5 shown.
[0081] (2) 80% of the standardized flood excerpt data was used as the training set of the model, and the other 20% was used as the validation set. The standardized water level, flow, rainfall, evaporation and temperature were used as the input data of the model, and the measured flow data of Hankou Station was used as the label training model;
[0082] (3) The evaluation indicators selected for model comparison analysis include Nash efficiency coefficient (NSE), determination coefficient (R 2 ), root mean square error (RMSE) and mean absolute error (MAE), using NSE and R 2 The accuracy of the model simulation results is evaluated. The closer the value is to 1, the better the simulation effect is. RMSE and MAE are used to evaluate the error of the model results. The smaller the value is, the smaller the error of the model results is. The specific formulas are:
[0083]
[0084] Where: Y i and are the real value and the simulated value respectively; Y i 'and are the average of the true value and the average of the simulated value, respectively.
[0085] The simulation results of the TCN-Markov model of the present invention, the traditional CNN prediction model and the LSTM prediction model are as follows: Figure 6-Figure 8 As shown in the figure, from a global perspective, compared with the traditional CNN forecast model and LSTM forecast model, the overall trend of the simulation results of the TCN-Markov model of the present invention is closer to the real flood process. From the perspective of local randomness, the simulated values of the key nodes of the flood process of the TCN-Markov model of the present invention are closer to the real values.
[0086] The results show that the Markov-constrained loss function has great reverse guidance significance in the training process of the TCN-Markov model of the present invention, so that the TCN-Markov model of the present invention can pay attention to the impact of random hydrological processes on simulation results during the training process.
[0087] The evaluation results are shown in Table 2. It can be seen from Table 2 that the NSE and R 2 The highest, indicating that the TCN-Markov model of the present invention has the highest simulation accuracy for the flood process; the RMSE and MAE of the TCN-Markov model of the present invention are the lowest, indicating that the error of the simulation result of the TCN-Markov model of the present invention is the smallest.
[0088] Table 2 Evaluation results of three models
[0089]
[0090] In summary, the present invention proposes a deep learning flood forecasting method considering random hydrological processes, that is, the loss function of the time domain convolutional neural network (TCN) model is improved according to the Markov state transition probability index, and the back propagation of the loss function in the deep learning model training process is used to guide the training model, enhance the model training's local dependence on the flood process and the ability to capture key nodes, and construct a TCN-Markov model; compared with the traditional CNN forecasting model and LSTM forecasting model, the TCN-Markov model considering random hydrological processes of the present invention has a better control over the global and key nodes of the flood process.
[0091] Example 2
[0092] The electronic device of the present invention comprises a central processing unit and a memory, wherein the central processing unit is used to call and run a computer program stored in the memory to execute the steps of the deep learning flood forecasting method considering random hydrological processes.
[0093] Example 3
[0094] The computer-readable storage medium of the present invention stores a computer program implemented according to the deep learning flood forecasting method considering random hydrological processes in the form of computer-readable instructions. When the computer program is called and executed by a computer, the steps included in the corresponding method are executed.
[0095] The above is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited to the above content. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.
Claims
1. A deep learning flood forecasting method considering random hydrological processes, characterized in that: The following steps are involved: Step S1: Collect the hydrological and meteorological data of the basin during the flood period, and standardize the collected hydrological and meteorological data; Step S2: Classify the flood process according to the hourly flow rate, and use the Markov model to construct the state transition probability index of different flood process levels; Step S3: Using the time domain convolutional neural network model as the basic model, the loss function in the time domain convolutional neural network model is improved by the state transition probability index, so as to construct a deep learning flood forecasting model (TCN-Markov) considering random hydrological processes; Step S4: dividing the standardized hydrological and meteorological data into a training set and a validation set, training the constructed deep learning flood forecasting model through the training set and validating the deep learning flood forecasting model through the validation set, to obtain a trained deep learning flood forecasting model; Step S5: Forecast floods using the trained deep learning flood forecasting model.
2. The deep learning flood forecasting method considering random hydrological processes according to claim 1 is characterized in that: In step S1, the hydrological and meteorological data of the flood period of the basin include five-dimensional data of hourly water level, hourly flow, rainfall, evaporation and temperature for many years. The water level and flow data directly reflect the flood process of the basin during the flood period, the rainfall and evaporation data reflect the indirect impact on the flood process of the basin from the water vapor circulation intensity, and the temperature data reflects the potential impact on the flood process of the basin.
3. The deep learning flood forecasting method considering random hydrological processes according to claim 2 is characterized in that: In step S1, the data standardization processing method selects the maximum minimum standardization method, and the specific formula is as follows: In the formula, x i 、x i ′ are the data before and after standardization, respectively, x min 、x max are the minimum and maximum values of the data series, respectively.
4. The deep learning flood forecasting method considering random hydrological processes according to claim 3 is characterized in that: In step S2, the flood process is classified according to the hourly flow, specifically into base flow period (L1), flood period (L2), flood peak period (L3), water receding period (L4) and recovery period (L5); at the same time, preconditions are added to assist in determining the flood process classification at a certain moment; wherein, The prerequisite for judging whether a flood process at a certain moment is the base flow period is that the previous period is the base flow period or the flood period; The precondition for judging whether a flood process at a certain moment is a flood period is that the previous period is a base flow period or a flood period; The prerequisite for judging whether a flood process at a certain moment is a receding period is that the previous period is a flood peak period or a receding period; The prerequisite for judging whether a flood process at a certain moment is a recovery period is that the previous period is a flood peak period, a receding period or a recovery period.
5. The deep learning flood forecasting method considering random hydrological processes according to claim 4 is characterized in that: In step S2, the Markov model is used to construct the state transition probability index of different flood process levels. The specific formula is: Where: P ij For state L i Transfer to state L j The probability of Q ij For state L i Transfer to state L j The frequency, Q i For state L i The frequency of occurrence; among them, The state transition probability index of the base flow period includes: 11 , base flow period turns to flood period P 12 , the rapid transition from base flow to peak flow caused by extreme rainfall 13 ; The probability index of state transition during the flood period includes: the transition from the flood period to the base flow period due to insufficient precipitation P 21 , the flood season turns to the flood seasonP 22 , the flood season turns to the peak seasonP 23 ; The state transition probability index during the flood peak period includes: the flood peak period turns to the flood peak period P 33 , flood peak period turns to receding periodP 34 , the flood peak period quickly turns to the recovery period due to human intervention in flood discharge P 35 ; The state transition probability index of the water-reduction period includes: 44 , the water withdrawal period turns to the recovery period P 45 ; The state transition probability index of the recovery period includes: 55 .
6. The deep learning flood forecasting method considering random hydrological processes according to claim 5 is characterized in that: In step S3, the time domain convolutional neural network model includes 3 residual blocks, 1 Flatten layer and 1 Dense fully connected layer, wherein each residual block contains 3 convolutional layers, and the number of neurons in the 3 convolutional layers is 128, 128 and 64 respectively; the number of neurons in the Dense fully connected layer is 64; the activation function is Relu, the optimizer is Adam, the learning rate is 0.001, and the loss rate is 0.
2.
7. The deep learning flood forecasting method considering random hydrological processes according to claim 6 is characterized in that: In step S3, the baseline loss function of the time-domain convolutional neural network model is: Where: Loss p is the benchmark loss function, Y i is the true value; is the simulated value, n is the number of samples; The benchmark loss function of the time-domain convolutional neural network is improved by the Markov state transition probability index. The specific improvement formula is: In the formula, Loss n is the loss function of the Markov constraint, Y i is the true value; is the simulation value, P i is the state transition probability index corresponding to the true value; P i ′ is the state transition probability index corresponding to the simulation value, and n is the number of samples.
8. The deep learning flood forecasting method considering random hydrological processes according to claim 1 is characterized in that: In step S4, 80% of the standardized hydrological and meteorological data are used as a training set, and the remaining 20% are used as a validation set.
9. An electronic device, comprising a central processing unit and a memory, characterized in that: The central processing unit is used to call and run the computer program stored in the memory to execute the steps of the deep learning flood forecasting method considering random hydrological processes as described in any one of claims 1 to 8.
10. A computer-readable storage medium, characterized in that: It stores a computer program implemented according to the deep learning flood forecasting method considering random hydrological processes as described in any one of claims 1 to 8 in the form of computer-readable instructions, and when the computer program is called and executed by a computer, the steps included in the corresponding method are executed.