A Neural Network Basin Rainfall-Runoff Forecasting Method and System Considering Initial Loss

Abstract

The present invention provides a neural network-based basin rainfall-runoff forecasting method and system considering initial losses, including: obtaining a rainfall-flood runoff time series based on historical rainfall and flood runoff data within the basin; calculating the initial rainfall loss, subtracting the initial rainfall loss from the rainfall-flood runoff time series to obtain potential runoff sequence data, and integrating the potential runoff sequence data and the rainfall-flood runoff time series to generate a gated recurrent unit training data set; constructing a GRU neural network model; setting hyperparameters of the GRU neural network model, training the GRU neural network model through the training data set to obtain the hidden mapping relationship between input and output, and obtaining the trained GRU neural network model; inputting real-time rainfall data and runoff into the trained GRU neural network model to perform runoff prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of basin rainfall-runoff forecasting, and in particular relates to a neural network basin rainfall-runoff forecasting method and system considering initial loss. Background Technique

[0002] The statements in this part only provide background technical information related to the present invention, and do not necessarily constitute prior art.

[0003] The inventors found that: at present, deep learning methods have been widely applied in the field of flood forecasting. However, existing research usually optimizes the structure of deep learning network models or couples them with other models to improve forecasting performance. Most models are old and lack consideration of rainfall-runoff mechanism factors, making it difficult to meet the requirements of accuracy and timeliness for flood forecasting of small and medium-sized rivers. Summary of the Invention

[0004] To overcome the deficiencies of the above-mentioned prior art, the present invention provides a neural network basin rainfall-runoff forecasting method considering initial loss, which can effectively improve the accuracy and speed of rainfall-runoff forecasting and flood evolution simulation.

[0005] To achieve the above object, one or more embodiments of the present invention provide the following technical solutions:

[0006] In the first aspect, a neural network basin rainfall-runoff forecasting method considering initial loss is disclosed, including:

[0007] Based on historical rainfall and flood runoff data in the basin, a rainfall-flood runoff time series is obtained;

[0008] The initial rainfall loss is obtained, and the potential runoff sequence data is obtained by deducting the initial rainfall loss from the rainfall-flood runoff time series. The potential runoff sequence data and the rainfall-flood runoff time series are integrated to generate a gated recurrent unit training dataset;

[0009] A GRU neural network model is constructed, with the potential runoff sequence data as the input variable of the GRU neural network model and the corresponding future-period runoff sequence data as the output variable of the GRU neural network model;

[0010] The hyperparameters of the GRU neural network model are set, and the GRU neural network model is trained through the training dataset to obtain the hidden mapping relationship between the input and output, and the trained GRU neural network model is obtained;

[0011] The real-time rainfall data and runoff are input into the trained GRU neural network model for runoff prediction.

[0012] As a further technical solution, the obtained rainfall-flood runoff time series is the sub-basin surface rainfall obtained by converting the control area of rainfall stations in the basin using the Thiessen polygon method and the flood runoff time series at the basin outlet.

[0013] As a further technical solution, the method for obtaining the initial loss is as follows:

[0014] I0 = f(d)

[0015] d = W H -V m

[0016] Where, I0 is the initial loss value, d is the soil moisture saturation deficit, W H is the soil saturation water content, and V m is the soil saturation water content before rainfall.

[0017] As a further technical solution, it also includes: performing data standardization processing on the generated training data set, and processing the original sample data set into a numerical value with a value range of [0, 1].

[0018] As a further technical solution, the steps for training the GRU neural network model are as follows:

[0019] The precipitation sequence of m rainfall stations in the basin for n time periods, and the potential runoff sequence data obtained by deducting rainfall losses from the precipitation, the two constitute an (m + 1) × n-dimensional input matrix, and the training target matrix is the predicted runoff at the t + 1 moment;

[0020] Determine the hyperparameters of the GRU network model, including: batch size, number of network layers, number of neurons, and type of optimizer;

[0021] In the forward propagation stage of model training, data information is transmitted forward from the input matrix to the output matrix, data feature values are extracted, and results are obtained;

[0022] When the result calculated by forward propagation does not meet the error requirement, enter the backpropagation stage, and distribute the error from the output to the input to different parameters for training and correction;

[0023] The model is continuously iterated until the accuracy requirement is met, and the model training is completed.

[0024] As a further technical solution, the real-time rainfall data and runoff are obtained in real time according to the basin water conservancy cloud platform database.

[0025] As a further technical solution, using the trained GRU neural network model for runoff prediction mainly includes the following steps:

[0026] Search for the data of rainfall stations and hydrological stations in the basin according to the basin water conservancy cloud platform database, and obtain the rainfall data and runoff data from the current moment T0 to the moment T within the flood forecasting period with a time span. 0-n The rainfall data and runoff data at the moment of T.

[0027] Obtain an accurate and reliable initial rainfall loss Ia.

[0028] Deduct the initial rainfall loss Ia from the rainfall data, and form a potential runoff-runoff matrix with the runoff data as the input data.

[0029] Input the input data into the trained GRU neural network to obtain the runoff value within the forecasting period.

[0030] In a second aspect, a neural network basin rainfall-runoff forecasting system considering initial loss is disclosed, including:

[0031] A training dataset acquisition module, configured to: based on the historical rainfall and flood runoff data in the basin, obtain a rainfall-flood runoff time series;

[0032] Obtain the initial rainfall loss, deduct the initial rainfall loss from the rainfall-flood runoff time series to obtain the potential runoff series data, and integrate the potential runoff series data and the rainfall-flood runoff time series to generate a gated recurrent unit training dataset;

[0033] A GRU neural network model construction and training module, configured to: construct a GRU neural network model, use the potential runoff series data as the input variable of the GRU neural network model, and the corresponding future period runoff series data as the output variable of the GRU neural network model;

[0034] Set the hyperparameters of the GRU neural network model, train the GRU neural network model through the training dataset to obtain the hidden mapping relationship between the input and output, and obtain the trained GRU neural network model;

[0035] A runoff prediction module, configured to: input the real-time rainfall data and runoff into the trained GRU neural network model for runoff prediction.

[0036] The above one or more technical solutions have the following beneficial effects:

[0037] The neural network basin rainfall-runoff forecasting method considering initial loss of the present invention uses a GRU neural network as a prediction tool, which has a faster convergence speed compared with the previous neural network, and better performance in parameter update and generalization.

[0038] The neural network-based rainfall-runoff forecasting method for river basins considering initial losses in the present invention takes into account more rainfall-runoff mechanistic factors. By combining with HEC-HMS, the deep learning network is improved to enhance the forecasting accuracy of deep learning, enrich the methods for river flood forecasting, and improve the accuracy and timeliness of flood simulation and forecasting.

[0039] Advantages of additional aspects of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings forming a part of this specification are used to provide a further understanding of the present invention. The schematic embodiments and descriptions thereof of the present invention are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0041] Figure 1 is the forecasting flowchart of the present disclosure;

[0042] Figure 2 is the sub-basin division map of the present disclosure;

[0043] Figure 3 is the model training process of the present disclosure;

[0044] Figure 4 is the model input-output structure of the present disclosure;

[0045] Figure 5 is the comparison diagram of forecasting results in the embodiment of the present disclosure;

[0046] Figure 6 is the comparison diagram of forecasting discharge in the embodiment of the present disclosure;

[0047] Figure 7 is the comparison diagram of forecasting discharge of the LSTM model and the Ia-GRU model in the test set;

[0048] Figure 8 is the schematic diagram of the model structure in the embodiment of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0050] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention.

[0051] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0052] Example 1

[0053] See the appendix Figure 1 As shown, this example discloses a neural network-based basin rainfall-runoff forecasting method considering initial loss. This method optimally selects the initial rainfall loss through the HEC-HMS (Hydrologic Engineering Center - Hydrologic Modeling System), subtracts the initial rainfall loss from the rainfall amount to obtain the potential runoff, constructs a GRU network-based basin rainfall-runoff forecasting model considering initial loss, and realizes the rapid forecasting of rainfall runoff.

[0054] This method includes the following steps:

[0055] (1) Data collection: Collect historical rainfall and flood runoff data in the basin over the years and organize them into corresponding rainfall-flood runoff time series.

[0056] The rainfall-flood runoff time series specifically is: the rainfall data series and the flood runoff series at the basin outlet corresponding to the corresponding time. Among them, the rainfall data series needs to be further processed into a potential runoff data series as the input for training the model; the flood runoff series is both the input and output of neural network training, that is, the runoff at the previous moment is the input, and the runoff at the next moment is the target output.

[0057] (2) Optimally select potential runoff and preliminarily construct a training dataset: Use the HEC-HMS hydrological model to optimally obtain accurate and reliable initial rainfall loss Ia based on data such as rainfall runoff, land use, soil type, and DEM digital elevation. On the basis of the initial rainfall loss Ia obtained by optimizing the HEC-HMS parameters, subtract the initial rainfall loss Ia from the rainfall data series to obtain the potential runoff series data, and integrate the potential runoff series data and the runoff series data to generate a GRU (Gate Recurrent Unit, GRU) training dataset.

[0058] (3) Dataset processing: Perform data standardization processing on the dataset and divide the dataset into a training set and a test set in a ratio of 7:3.

[0059] (4) Model training: Use the potential runoff-runoff time series as the network input variable, and the corresponding future period runoff series data as the output variable to construct a GRU neural network model, set the hyperparameters of the network, and obtain the hidden mapping relationship between the input and output through training.

[0060] The above-mentioned potential runoff-runoff time series refers to the potential runoff and the flood runoff series at the outlet of the basin corresponding to the time.

[0061] (5) Flood prediction: According to the real-time rainfall data and runoff in the basin water conservancy cloud platform database, the trained GRU neural network model is used to predict the runoff.

[0062] The present invention uses a GRU network to construct a basin rainfall-runoff forecasting model, considers the runoff generation and concentration process of the rainfall process based on the initial loss and subsequent loss method, and uses the potential runoff as an input variable, effectively improving the accuracy and speed of rainfall-runoff forecasting and flood routing simulation.

[0063] In this embodiment, the rainfall-flood runoff data series in step 1 is the sub-basin surface rainfall obtained by converting the control area of the rainfall stations in the basin using the Thiessen polygon method and the flood runoff time series at the basin outlet.

[0064] As Figure 2 shown, where W60 to W100 are the sub-basins divided. The study area is divided using the Thiessen polygon. Each Thiessen polygon partition after division contains a rainfall station, that is, it is considered that the rainfall in the Thiessen polygon area is contributed by this rainfall station. Then the sub-basin rainfall is the rainfall contribution of each rainfall station Thiessen polygon partition in the proportion of the area of this sub-basin.

[0065]

[0066] Among them, P is the sub-basin rainfall, p i represents the rainfall at the i-th rainfall station in this sub-basin, A i ' represents the area of the Thiessen polygon partition of the i-th rainfall station in this sub-basin, A i is the total area of the Thiessen polygon partition of the i-th rainfall station.

[0067] In this embodiment, in step 2, the initial rainfall loss is obtained based on the initial loss and subsequent loss method according to the following formula:

[0068] I0 = f(d)

[0069] d = W H -V m

[0070] Among them, I0 is the initial loss value, d is the soil moisture saturation difference, W H is the soil saturation water content, V m is the soil saturation water content before rainfall.

[0071] After that, the potential runoff is obtained based on the initial loss. The potential runoff is the rainfall remaining after deducting losses such as infiltration, vegetation interception, and evaporation. The expression is as follows:

[0072]

[0073] Wherein is the potential runoff, pt is the rainfall, and Ia is the initial loss, all in mm.

[0074] In this embodiment, the initial loss should be the total loss of water before runoff occurs, that is, it does not include the infiltration loss after runoff.

[0075] In this embodiment, the normalization method in step 3 is deviation normalization. After processing the original sample data set, it becomes a value in the range of [0,1]. The conversion function is:

[0076]

[0077] Wherein, x * represents the normalized value, x represents the original sample data, x max represents the maximum value of the sample data, and x min represents the minimum value in the sample data.

[0078] In this embodiment, the specific steps of training the GRU neural network model in step 4 are

[0079] (4-1) The precipitation sequences (p t-n+1 , p t-n+2 , p t-n+3 ,..., p t ) of m rain gauges in the basin for n time periods, and the potential runoff sequence data obtained by deducting the rainfall loss from the precipitation These two constitute an (m + 1) × n-dimensional input matrix, and the training target matrix is the predicted runoff Q at time t + 1 n+1 ;

[0080] (4-2) Determine the hyperparameters of the GRU network model, including: batch size, number of network layers, number of neurons, type of optimizer, etc.;

[0081] (4-3) In the forward propagation stage of model training, data information is transmitted forward from the input matrix to the output matrix, data feature values are extracted, and results are obtained;

[0082] The so-called data feature values of the GRU network generally refer to time series feature values, that is, the feature h at the previous moment t-1 will affect the matrix operation at time t;

[0083] (4-4) When the result of the forward propagation calculation does not meet the error requirement, enter the backpropagation stage, and distribute the error from the output to the input to different parameters for training and correction;

[0084] (4-5) The model is continuously iterated until the accuracy requirement is met, and the model training is completed.

[0085] Specifically, when training the network model, the objective function is as follows

[0086]

[0087] where n is the number of dataset samples, Y i is the target true runoff, and f(x i ) is the predicted value of the model. During the training process, iterative training is carried out with the aim of minimizing the objective function until the maximum number of iterations is reached or the network learning rate converges.

[0088] As Figure 8 shown is the GRU neural network structure. The network receives the current input x t and the hidden state h t-1 from the previous calculation. The GRU outputs y t and passes the current hidden state h t . The specific calculation steps are as follows:

[0089] (2-1) After receiving the input, z is the update gate, which determines which past information is retained for the future; r is the reset gate, which mainly determines which information is forgotten. The expressions are as follows respectively:

[0090] z = σ(W (z) x i + U (z) h i-1 ) (2)

[0091] r = σ(W (r) x i + U (r) h i-1 ) (3)

[0092] where σ is the Sigmoid activation function, W (z) , U (z) , W (r) , U (r) are their respective weight matrices, and the meanings of other symbols are the same as those described above. The update gate and the reset gate jointly determine which information is finally used as the network output. Their special feature is that they can preserve the information in the long-term sequence and will not disappear over time or be removed because they are not relevant to the prediction result.

[0093] (2-2) After obtaining the gated information, the "reset" is obtained and then h t-1 ' and x tConcatenation, activated by the tanh function, is as follows:

[0094]

[0095] Among them, W and U are weight matrices, and the meanings of other symbols are the same as those described above. Here, h' contains the current input x t information, which is equivalent to memorizing the current state.

[0096] (2-3) In this way, the network performs two steps of forgetting and memorizing, using z to update the memory h t , and the expression is as follows:

[0097]

[0098] (2-4) The calculation formula for the final output value is:

[0099] y t = σ(W y h t + B0) (6)

[0100] Among them, W y is the weight matrix, B0 is the output bias term, and the meanings of other symbols are the same as those described above.

[0101] It can be seen from the above mechanism that the advantage of the GRU network is that it uses a gating unit z to perform forgetting and selective memorization at the same time. This structural design reduces the training cost and greatly improves the convergence efficiency of the neural network.

[0102] In this embodiment, the role of the GRU network is to replace the post-loss calculation and confluence calculation processes in the traditional hydrological model. Due to the characteristics of flood forecasting in small and medium-sized basins, the uncertainty of post-loss and confluence calculations in the basin is more than that of runoff generation calculations, which greatly reduces the accuracy of flood forecasting. Based on this consideration, the hydrological model is combined with the GRU network for rainfall-runoff forecasting.

[0103] In this embodiment, the step 5 of using the trained GRU neural network model to predict the runoff according to the real-time rainfall data and runoff in the basin water conservancy cloud platform database mainly includes the following steps:

[0104] (5-1) Search for the data of rain gauges and hydrological stations in the basin in the basin water conservancy cloud platform database to obtain the rainfall data and runoff data from the current time T0 to the time T at the flood forecasting period with a time span; 0-n moment;

[0105] (5-2) Use HEC-HMS optimization to obtain an accurate and reliable initial rainfall loss Ia;

[0106] (5-3) Deduct the initial rainfall loss Ia from the rainfall data, and form a potential runoff-runoff matrix with the runoff data as the input data;

[0107] (5-4) Input the input data into the trained GRU neural network to obtain the runoff value within the prediction period.

[0108] The following further elaborates on the traceability method of the present disclosure with examples.

[0109] (1) Data collection:

[0110] The study area is the Yufu River Basin in Jinan City, Shandong Province. The upper reaches of the Yufu River Basin belong to the temperate continental monsoon climate, with significant monsoon characteristics. Its main climate features are: obvious monsoon, distinct seasons; cold in winter and hot in summer, with concentrated rainfall. Precipitation is concentrated during the flood season, with uneven spatial and temporal distribution, and local heavy rains occur from time to time, with strong suddenness. The average annual precipitation in the basin is 686.3 mm, the maximum annual precipitation (in 1964) is 1058 mm, the minimum annual precipitation (in 2002) is 370 mm, and the interannual and intra-annual distributions are extremely unbalanced. The average precipitation in the northern plain area is 650 - 700 mm, and slightly more in the southern mountainous area, which is 700 - 800 mm. Due to the influence of the monsoon climate, the seasonal distribution of precipitation is uneven, dry and less rainy in spring, hot and rainy in summer, and the precipitation during the flood season can account for about 70% of the annual precipitation, which is prone to form floods within the basin. There is also uneven distribution in the rainfall space.

[0111] Sub-basin data needs to be converted into the surface rainfall of the sub-basin by combining the control area of each rain gauge, and the Thiessen polygon method is used to determine the weight of the rain gauge in each sub-basin. The sub-basins are divided as Figure 2 shown, and the proportion of each rain gauge in the sub-basin area is as shown in Table 1 below:

[0112] Table 1 Proportion of each rain gauge in the sub-basin area

[0113]

[0114] (2) Optimally select the potential runoff and preliminarily construct the training data set

[0115] Based on data such as soil and land use types and DEM elevation data within the basin, construct the HEC-HMS model, and the optimized selection of the initial loss results is as follows:

[0116] Table 2

[0117]

[0118] (3) Data set processing

[0119] In this study, the flow data of hydrological stations in the Yufu River Basin from 1973 to 2020, as well as the long-term rainfall data during the flood season of 7 rain gauges, were collected. By analyzing the rainfall-runoff data and excluding flood events with poor representativeness, missing rainfall or runoff data, interpolation processing was performed on the rainfall and flow data. Finally, 30 flood processes with a total of 3,166 time periods of data were selected and divided into a training set and a test set. The input-output structure of the model is as shown in Figure 4 shown.

[0120] In a specific implementation example, the model data processing process is as follows:

[0121] (3-1) Historical flood flow data of the Yufu River Basin from 1973 to 2020, and multi-year flood season rainfall data of 7 hydrological stations and rain gauges were collected. Through data analysis and collation, flood events with complete rainfall or runoff data and strong representativeness were selected. Outlier correction and data supplementation were performed on the flow data and rainfall data. Finally, 30 flood processes with a time interval of 2 hours and a total of 3,068 time periods of data were selected to generate a data set, which was divided into a training set and a test set.

[0122] (3-2) The data of sub-basins need to be combined with the control area of each rain gauge to convert the areal rainfall of the sub-basins. The Thiessen polygon method is used to determine the weights of rain gauges in each sub-basin.

[0123] Table 3 Proportion of each rain gauge in the sub-basin area

[0124]

[0125] (3-3) The HEC-HMS model was used to simulate and optimize all flood events in the Yufu River Basin to obtain the initial rainfall loss Ia of the model driver. The obtained initial rainfall loss Ia will be applied to the processing of model input parameters.

[0126] Table 4 Optimization of initial rainfall loss I for different flood events a Table (unit: mm)

[0127]

[0128]

[0129] Table 5 Optimization of initial rainfall loss I for different flood events a Table (unit: mm)

[0130]

[0131] (3-4) The rainfall data needs to be processed with the initial infiltration loss Ia data to obtain the data set of the basin rainfall runoff model. Each rain gauge deducts the corresponding rainfall initial loss according to its sub-basin. The operation process is shown in the example in the table below:

[0132] Table 6 Model input data processing example (Xiying Station 20050918)

[0133]

[0134] (3-4) The net rainfall data and runoff data are sorted and organized into a data set according to the time series. The data are linearly transformed using deviation normalization. The original data are processed into numbers with values in the range of [0,1] and imported into the GRU network model for training and learning to explore the mapping relationship.

[0135] (4) Model training

[0136] The model training process is as follows Figure 3 shown.

[0137] The final hyperparameters selected were as follows: 2 hidden layers, 10 time steps, 256 neurons, and a batch size of 32. The model was trained using the RMSprop optimizer with a decay coefficient of 0.8, a learning rate of 0.0001, and 1000 iterations. The mean squared error (MSE) was used as the error function, which is the expected value of the square of the difference between the model's predicted value and the measured value.

[0138] (5) Flood forecast

[0139] Flood forecast results are available at Figure 5 .

[0140] In order to quantitatively evaluate the accuracy of rainfall-runoff simulation forecasts of the two models, four forecast error indicators, namely, peak flow relative error, peak time error, Nash coefficient, and correlation coefficient, are used to evaluate the accuracy of the two models in flood forecast simulation. They are peak flow relative error REP, peak time TEP, Nash coefficient NSE, and correlation coefficient r:

[0141]

[0142]

[0143]

[0144]

[0145] In the formula Indicates the flood peak forecast. represents the measured flood peak, Denotes the predicted peak arrival time, Denotes the measured peak arrival time, and Denote the predicted flow rate and the measured flow rate at time t respectively, Denotes the mean value of the measured flow rate, and n denotes the number of periods of flood events, Denotes and Covariance of and Denote respectively and Variances of

[0146] Table 7 Comparison table of predicted peak REP of LSTM model and Ia-GRU model

[0147]

[0148] Combined comparison of predicted flood peak and measured flood peak Figure 6 , In the prediction of flood peak flow rate, the relative error of the LSTM model without considering the initial loss is relatively large. Among them, the relative errors of the two flood events on September 18, 2005 and September 18, 2013 exceed 20%, while the Ia-GRU model is significantly reduced, and all three flood events are within the 20% permissible range. In contrast, the Ia-GRU model performs the most outstandingly in the simulation of the peak value, and the simulation effect of the flood peak flow rate of the flood on August 15, 2019 is the best.

[0149] Table 8 Comparison table of predicted peak arrival time TEP of LSTM model and Ia-GRU model

[0150]

[0151] According to Table 8 above, in the process of predicting rainfall-runoff, the LSTM model shows a lag in the peak arrival time. For the flood event on August 15, 2019, there is an error of 2 calculation periods, exceeding the permissible error of the peak arrival time. For the flood events on August 15, 1995 and September 18, 2013, there is an error of 1 calculation period. The reason may be that the peak flow rate increases rapidly during the rainfall-runoff process, increasing the training difficulty of the neural network and failing to accurately predict the peak arrival time. While the Ia-GRU performs better in terms of the peak arrival time of the model. Except for the flood event on September 18, 2005 where the predicted flood peak arrives one period earlier, the flood peaks of the other two flood events are accurately predicted.

[0152] Table 9 Comparison table of Nash coefficient NSE and correlation coefficient r of LSTM and Ia-GRU models

[0153]

[0154] The simulation effects of the two models for the flood of the 20000809 event are the best, with Nash coefficients all higher than 0.7 and correlation coefficients all higher than 0.9. The possible reason is that the runoff process of this event is relatively smooth, with only a single peak maximum flow value, resulting in better simulation and prediction effects. The simulation effect for the flood of the 20190815 event is the worst. The reason may be that the length of the flood sequence of this event is short, and the flow rate rises rapidly with two consecutive peaks, making it difficult for the neural network to simulate well. Combining the comparison chart of the predicted flow rates of the test sets of the LSTM model and the Ia-GRU model ( Figure 7 ), the Nash coefficients of the overall prediction results of the test sets are all above 0.8, and the correlation coefficients are all greater than 0.725. The degree of coincidence between the predicted flow rate and the measured flow rate shows a better effect, with a higher fitting degree.

[0155] In summary, from the three evaluation criteria of peak appearance time, Nash coefficient, and relative error, this prediction method is very accurate for rainfall-runoff prediction.

[0156] Embodiment 2

[0157] The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the above method are implemented.

[0158] Embodiment 3

[0159] The purpose of this embodiment is to provide a computer-readable storage medium.

[0160] A computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps of the above method are executed.

[0161] Embodiment 4

[0162] The purpose of this embodiment is to provide a neural network basin rainfall-runoff prediction system considering initial loss, including:

[0163] A training data set acquisition module, configured to: based on the historical rainfall and flood runoff data within the basin, obtain a rainfall-flood runoff time series;

[0164] Calculate the initial rainfall loss, subtract the initial rainfall loss from the rainfall-flood runoff time series to obtain potential runoff sequence data, and integrate the potential runoff sequence data and the rainfall-flood runoff time series to generate a gated recurrent unit training data set;

[0165] The GRU neural network model construction and training module is configured to: construct a GRU neural network model, use the potential runoff series data as the input variable of the GRU neural network model, and the corresponding future-period runoff series data as the output variable of the GRU neural network model;

[0166] Set the hyperparameters of the GRU neural network model, train the GRU neural network model through the training dataset to obtain the hidden mapping relationship between the input and output, and obtain the trained GRU neural network model;

[0167] The runoff prediction module is configured to: input the real-time rainfall data and runoff into the trained GRU neural network model for runoff prediction.

[0168] In the devices of the above second, third, and fourth embodiments, the steps involved correspond to those of the method embodiment one. For the specific implementation manners, reference may be made to the relevant description part of the first embodiment. The term "computer-readable storage medium" should be understood to include a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0169] Those skilled in the art should understand that the above-mentioned modules or steps of the present invention can be implemented by a general computer device. Optionally, they can be implemented by program codes executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0170] Although the specific implementation manners of the present invention have been described above in conjunction with the drawings, it is not a limitation to the protection scope of the present invention. Those skilled in the art should understand that based on the technical solutions of the present invention, various modifications or deformations that can be made without creative efforts by those skilled in the art are still within the protection scope of the present invention.

Claims

1. A neural network-based rainfall-runoff forecasting method for river basins considering initial losses, characterized in that, Including: Based on the historical rainfall and flood runoff data within the basin, obtain the rainfall-flood runoff time series; Calculate the initial rainfall loss, subtract the initial rainfall loss from the rainfall-flood runoff time series to obtain the potential runoff series data, and integrate the potential runoff series data and the rainfall-flood runoff time series to generate the gated recurrent unit training dataset; Construct a GRU neural network model, use the potential runoff series data as the input variable of the GRU neural network model, and the corresponding future period runoff series data as the output variable of the GRU neural network model; Set the hyperparameters of the GRU neural network model, and train the GRU neural network model through the training dataset to obtain the hidden mapping relationship between the input and output, and obtain the trained GRU neural network model; Input the real-time rainfall data and runoff into the trained GRU neural network model to predict the runoff; Using the trained GRU neural network model for runoff prediction includes the following steps: Search for the rainfall station and hydrological station data within the basin in the basin water conservancy cloud platform database, and obtain the rainfall data and runoff data from the current moment T0 to the moment T within the flood forecast period with a time span. 0-n Rainfall data and runoff data at the moment. Obtaining accurate and reliable initial rainfall losses Ia ; Deduct the initial rainfall loss from the rainfall data Ia and form a potential runoff-runoff matrix with the runoff data as the input data; Input the input data into the trained GRU neural network to obtain the runoff value within the forecast period.

2. The neural network-based watershed rainfall-runoff forecasting method considering initial losses as described in claim 1, wherein, The obtained rainfall-flood runoff time series is the sub-basin surface rainfall converted by using the Thiessen polygon method to calculate the control area of the rain gauges within the basin and the flood runoff time series at the basin outlet.

3. The neural network-based basin rainfall-runoff forecasting method considering initial losses as described in claim 1, characterized in that, The method for calculating the initial loss is: Among them, is the initial loss value, d is the difference in soil moisture saturation, is the saturated soil water content, is the saturated soil water content before rainfall.

4. A neural network-based rainfall-runoff forecasting method for river basins considering initial losses as claimed in claim 1, characterized in that, Also including: Perform data standardization processing on the generated training dataset, and process the original sample dataset into a value within [0, 1].

5. The neural network-based basin rainfall-runoff forecasting method considering initial losses as claimed in claim 1, characterized in that The steps for training the GRU neural network model are: Precipitation sequences of m rain gauges in the basin over n time periods, and potential runoff sequence data obtained by deducting rainfall losses from precipitation. These two form an (m + 1)×n dimensional input matrix, and the training target matrix is the predicted runoff at Formulate the hyperparameters of the GRU network model, including: batch size, number of network layers, number of neurons, and type of optimizer; In the forward propagation stage of model training, transfer data information from the input matrix to the output matrix in the forward direction, extract data feature values, and obtain the result; When the result calculated by the forward propagation does not meet the error requirement, perform the backward propagation stage, and distribute the error from the output to the input to different parameters for training and correction; The model is continuously iterated and repeated until the accuracy requirement is met, and the model training is completed.

6. A neural network-based basin rainfall-runoff forecasting method considering initial losses as claimed in claim 1, characterized in that, The real-time rainfall data and runoff are obtained in real time according to the basin water conservancy cloud platform database.

7. A neural network-based rainfall-runoff forecasting system for river basins considering initial losses, characterized in that, Including: Training dataset acquisition module, configured to: Based on the historical rainfall and flood runoff data within the basin, obtain the rainfall-flood runoff time series; Calculate the initial rainfall loss, subtract the initial rainfall loss from the rainfall-flood runoff time series to obtain the potential runoff series data, and integrate the potential runoff series data and the rainfall-flood runoff time series to generate the gated recurrent unit training dataset; GRU neural network model construction and training module, configured to: Construct a GRU neural network model, use the potential runoff series data as the input variable of the GRU neural network model, and the corresponding future period runoff series data as the output variable of the GRU neural network model; Set the hyperparameters of the GRU neural network model, and train the GRU neural network model through the training dataset to obtain the hidden mapping relationship between the input and output, and obtain the trained GRU neural network model; The runoff prediction module is configured to: input real-time rainfall data and runoff into the trained GRU neural network model for runoff prediction; Using the trained GRU neural network model for runoff prediction includes the following steps: Search for the rainfall station and hydrological station data within the basin in the basin water conservancy cloud platform database to obtain the rainfall data and runoff data from the current moment T0 to the moment T within the flood forecast period with a time span. 0-n Rainfall data and runoff data at the moment; Obtain accurate and reliable initial rainfall losses Ia ; Deduct the initial rainfall loss from the rainfall data Ia , and form a potential runoff - runoff matrix with the runoff data as the input data; Input the input data into the trained GRU neural network to obtain the runoff value within the prediction period.

8. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1-6 above.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it executes the steps of the method according to any one of claims 1-6 above.

Images