An intelligent forecasting method for flood processes in mountainous areas with insufficient data based on LSTM-SCS

Abstract

The present invention discloses an intelligent forecasting method for flood processes in mountainous areas with insufficient data based on LSTM-SCS. The method comprises the following steps in sequence: collecting and processing historical rainstorm monitoring data and obtaining underlying surface data at observation stations in mountainous areas with insufficient data, establishing an SCS model and completing flood process lines in the watershed areas of observation stations in mountainous areas with insufficient data, preprocessing input data of an LSTM neural network prediction model, establishing the LSTM neural network prediction model and evaluating prediction results. The SCS model established in the intelligent forecasting method for flood processes in mountainous areas with insufficient data based on LSTM-SCS has the characteristics of significant interpolation effect for insufficient data and high forecast accuracy.

Description

Technical Field

[0001] The present invention relates to an intelligent forecasting method for flood processes in mountainous areas with insufficient data based on LSTM-SCS. Background Art

[0002] The rapid flow, high volume, and sudden rise and fall of floods in mountainous areas contribute to their sudden and destructive nature. Furthermore, the complex conditions of small watersheds in mountainous areas, coupled with numerous influencing factors, coupled with underdeveloped hydrological observation infrastructure and a lack of hydrological observation data, complicate hydrological forecasting in mountainous areas. The scarcity of ground-based observations in western and eastern mountainous areas of my country severely limits the accuracy of flood forecasts in these regions. Currently, mountain flood forecasting methods can be summarized into three categories: flash flood process forecasting based on distributed hydrological models (referred to as distributed hydrological models), flash flood warnings based on critical rainfall thresholds, and artificial intelligence forecasting. Flash flood warnings based on critical rainfall thresholds rely on obtaining critical rainfall in small watersheds through water level / flow inversion, analogy, and hydrodynamic methods. However, these methods lack description and analysis of runoff generation mechanisms and cannot fully capture the entire flash flood disaster process. Distributed hydrological models account for the spatial variability of precipitation and underlying surfaces, making better use of spatial information captured by geographic information systems (GIS) and remote sensing technologies to describe the process from rainfall to runoff. However, distributed hydrological models divide the watershed into many response units to reflect spatiotemporal variability. The internal parameters of these units vary, inevitably leading to errors. Artificial intelligence forecasting is an emerging method, and machine learning is a fundamental implementation of artificial intelligence. It seeks input-output relationships from data to better fit nonlinear runoff sequences. However, the drawback is that it requires a large training dataset. Generally, data from small watersheds in mountainous areas is scarce, insufficient to support model training. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide an intelligent forecasting method for flood processes in mountainous areas with insufficient data based on LSTM-SCS.

[0004] The technical solution of the present invention is to realize the intelligent forecasting method of flood process in mountainous area with insufficient data based on LSTM-SCS. The intelligent forecasting method of flood process in mountainous area with insufficient data based on LSTM-SCS comprises the following steps in sequence:

[0005] (1) Collection and processing of historical flood and rainstorm data in the basin area of ​​mountainous observation stations with insufficient data and acquisition of underlying surface data

[0006] ①. Collection and processing of historical flood and rainstorm data in the watershed of the observation station in the mountainous area with lack of data: collect historical flood and rainstorm data of at least 12 floods that lasted for at least 24 hours in the watershed of the observation station in the mountainous area with lack of data, and the historical flood and rainstorm data include historical rainfall data, historical flood peak flow and historical peak time; if there is no rainfall measured data in the watershed of the observation station in the mountainous area with lack of data, the rainfall measured data in the watershed of the surrounding observation station is used for interpolation to obtain historical rainfall data of at least 12 floods that lasted for at least 24 hours in the watershed of the observation station in the mountainous area with lack of data at time intervals of hours and arranged in chronological order; according to the data of the observation station in the mountainous area with lack of data The historical peak flows of at least 12 floods lasting at least 24 hours in the watershed of the data-deficient mountain observation station are calculated based on flood traces in the watershed of the data-deficient mountain observation station; the historical peak times of at least 12 floods lasting at least 24 hours in the watershed of the data-deficient mountain observation station are determined based on historical records; the DEM data of the watershed of the data-deficient mountain observation station are subjected to depression filling and flow direction analysis based on ArcGIS, and the accumulated runoff is analyzed, the river network is extracted and classified, and the watershed basin is segmented to obtain calculation layer analysis data of the watershed basin and river system, and raster river network data is established based on the calculation layer analysis data;

[0007] ② Obtaining underlying surface data for watersheds of observation stations in mountainous areas with insufficient data: Based on the data on engineering facilities, topography, hydrology and meteorology, vegetation cover, river sections and soil types in the watersheds of observation stations in mountainous areas with insufficient data, the underlying surface data for the watersheds of observation stations in mountainous areas with insufficient data are obtained;

[0008] (2) Establishment of SCS model and completion of flood process lines in the basin area of ​​mountainous area with missing data

[0009] ①. Establishment of SCS model: a. Establish an SCS model for the watershed area of ​​the adjacent mountainous observation station with data based on the underlying surface data of the watershed area of ​​the adjacent mountainous observation station with data; b. Rate and adjust the parameters to obtain a suitable CN value and unit line based on the historical flow data and historical rainfall data of at least 12 floods with time intervals of hours and arranged in chronological order for at least 24 hours in the watershed area of ​​the adjacent mountainous observation station with data; c. Combine the CN value and unit line described in step (2) ①b with the underlying surface data of the watershed area of ​​the mountainous observation station with insufficient data in step (1) ② to construct an SCS model for the watershed area of ​​the mountainous observation station with insufficient data; d. Rate the CN value and unit line described in step (2) ①b based on the historical flood peak flow and historical peak time of the watershed area of ​​the mountainous observation station with insufficient data in step (1) ①;

[0010] ②. Completion of flood process lines in the watershed area of ​​the mountainous observation station with missing data: Based on the SCS model of the watershed area of ​​the mountainous observation station with missing data in step (2)① and the historical rainfall data of the watershed area of ​​the mountainous observation station with missing data in step (1)①, the flood process lines of the historical peak flow and historical peak time of at least 12 floods lasting at least 24 hours in the watershed area of ​​the mountainous observation station with missing data in step (1)① are completed to obtain the historical flow data of at least 12 floods lasting at least 24 hours in the watershed area of ​​the mountainous observation station with missing data, which are arranged in chronological order with time intervals of hours;

[0011] (3) Preprocessing of input data of LSTM neural network prediction model

[0012] Step (1) ① The seven rainfall data P1-P7 of each flood rainfall data in the historical rainfall data of the watershed area of ​​the data-deficient mountain observation station are arranged in chronological order to form a single rainfall data group for each flood, and the P7 moment of each single rainfall data group for each flood has a certain flow data Q of each flood flow data in the historical flow data corresponding to step (2) ②; the certain flow data Q corresponding to the P7 moment of each single rainfall data group for each flood and P1-P7 in each single rainfall data group for each flood are arranged in sequence to form a single rainfall data group for each flood. Flow data group; at least 18 of the single rainfall data groups of each flood or the single rainfall flow data groups of each flood are arranged in sequence according to the time of P1 in the data group to form each flood rainfall data group or each flood rainfall flow data group; the single rainfall data groups of each flood sorted in the each flood rainfall flow data group constitute each flood time series data set, and each data in the each flood time series data set is scaled to (0, 1) using max-min normalization as input data for each flood LSTM neural network prediction model;

[0013] (4) Establishment of LSTM neural network prediction model

[0014] The parameters of the LSTM neural network model are preliminarily formulated, including the number of hidden layers, the number of hidden layer nodes, and the number of epoch iterations. The forward propagation calculation formula of the Dropout neural network is added to the standard LSTM neural network as shown in formulas (1)-(4). 80% of the total number of floods in the training set and validation set after the preprocessing of the input data of the LSTM neural network prediction model in step (3) is used as the training set, and 20% as the validation set. The training set is input into the LSTM model for iterative training, and the training results are verified using the validation set until the prediction error meets the preset standard, and the prediction error sequence is output.

[0015] M(l)~Bernoulli(p)(1)

[0016] y ～(l) =r(l)*x(l)(2)

[0017] z i (l+1) =w i (l+1) y (～l) +b i (l+1) (3)

[0018] y i (l+1) =f(z i (l+1) )(4)

[0019] In formulas (1)-(4), M is a random mask matrix; l represents the layer index of the neural network, and the lth layer represents the current layer; for a neural network, the layers are arranged in order, from the input layer (usually denoted as the 0th layer) to the output layer; p is the retention rate; y ～(l) represents the output of layer l after the dropout operation; r(l) represents the dropout mask of layer l, usually represented as a vector of 0 or 1; x(l) represents the input of layer l; z i (l+1) represents the linear combination input of the i-th neuron in the l+1th layer; w i (l+1) is the weight vector connecting the i-th neuron in the l-th layer and the l+1-th layer; b i (l+1) represents the bias of the i-th neuron in the l-th layer; y i (l+1) represents the activation output of the i-th neuron in the l+1th layer; f is the activation function of the gate;

[0020] (5) Evaluation of prediction results

[0021] The flood process prediction results of the LSTM-SCS neural network are compared and analyzed with those of the BPNN-SCS neural network and the GRU-SCS neural network using the NSE and R indicators. The higher the values ​​of the NSE and R indicators, the higher the accuracy of the prediction results.

[0022] The data-deficient mountainous areas mentioned in the present invention refer to mountainous areas where historical flood and rainstorm data are easy to obtain, but historical flow data are difficult to obtain; the data-available mountainous areas mentioned in the present invention refer to mountainous areas with historical rainfall data and historical flow data. Unless otherwise specified, the historical rainfall data mentioned in the present invention refer to historical rainfall data of at least 12 floods (13 in the embodiment) that lasted for at least 24 hours, arranged in chronological order at time intervals of hours; unless otherwise specified, the historical flow data mentioned in the present invention refer to historical flow data of at least 12 floods (13 in the embodiment) that lasted for at least 24 hours, arranged in chronological order at time intervals of hours; unless otherwise specified, the historical peak flow rate mentioned in the present invention refers to the historical peak flow rate of at least 12 floods (13 in the embodiment) that lasted for at least 24 hours; unless otherwise specified, the historical peak appearance time mentioned in the present invention refers to the historical peak appearance time of at least 12 floods (13 in the embodiment) that lasted for at least 24 hours.

[0023] The SCS model described in step (2) of the method for intelligent forecasting of flood processes in mountainous areas with insufficient data based on LSTM-SCS is a flood model designed by the U.S. Soil Conservation Service for small watershed research. The SCS model has a simple structure and relatively few parameters. The SCS model is used to supplement hydrological data. For flood data that lacks actual flood process measurements and only has peak flow and peak time, the flood process line can be corrected based on the SCS model to obtain actual flood data as an input data set for intelligent forecasting of floods in small and medium-sized watersheds in mountainous areas. The SCS model is established based on the water balance equation and the two assumptions of equal proportions and the relationship between the maximum potential retention of initial loss values. Its formula is shown in formulas (5)-(6):

[0024]

[0025] In formulas (5)-(6): R is the surface runoff; P is the total rainfall; I a is the initial loss, i.e., the rainfall loss before surface runoff is generated, such as filling and vegetation interception; S is the maximum possible retention of the basin at that time, which is the upper limit of the subsequent loss; F is the actual subsequent loss;

[0026] The maximum possible soil retention capacity S is unevenly distributed in space and is related to data such as soil type, land use type, management strategy, and slope. It also changes with time as soil water content changes. The calculation formula for S in the SCS model is shown in formula (7):

[0027]

[0028] In formula (7): CN value is dimensionless and is a comprehensive parameter that reflects the basin characteristics before rainfall, integrating the factors of soil moisture, land use pattern and soil type condition before rainfall;

[0029] The data representation of the single rainfall flow data set of each flood event in step (3) is shown in Table 1:

[0030] Table 1 Data representation of single rainfall flow data set for each flood event

[0031]

[0032] In Table 1, Var1(t) and Var8(t) represent the time P7; Var2(t-1), Var3(t-2), Var4(t-3), Var4(t-4), Var4(t-5) and Var4(t-6) represent 6 hours, 5 hours, 4 hours, 3 hours, 2 hours and 1 hour before the time P7, respectively.

[0033] In step (4), the hydrological characteristics of the various small and medium-sized river basins in the Zhongshan area are similar. When conducting intelligent flood forecasting in each river basin, the parameter settings of each small and medium-sized river basin can refer to the same scheme; the three gates of the LSTM neuron in step (4) are the forget gate, input gate, and output gate. The forget gate controls whether to forget, and controls whether the hidden cell state of the previous layer will be forgotten in the LSTM with a certain probability, and selects the forgotten information; the input gate processes the current sequence position as input. The forget gate formula is shown in formula (8):

[0034] f t =σ(W f ·[h t-1 , x t ]+b f ) (8)

[0035] In formula (8), f t is the information that is forgotten in the sequence at time t; W f is the coefficient weight matrix of the forget gate; σ is the RELU activation coefficient; h t-1 is the hidden state output of the previous sequence LSTM module; x t represents the sample sequence input at time t; b f is the offset term of the forget gate;

[0036] The input gate formula is shown in formulas (9)-(10):

[0037] i t =σ(W i ·[h t-1 , x t ]+b i (9)

[0038]

[0039] The cell state (long-term) equation at time t is shown in formula (11):

[0040]

[0041] In formulas (9)-(11): i t is the activation value of the input gate; σ is the Sigmoid activation function; W i is the weight matrix of the input gate; h t-1 is the hidden state of the previous time step; x t is the input of the current time step; b i is the bias vector of the input gate; C ～ t is the candidate memory unit state; tanh is the hyperbolic tangent activation function; W c is the weight matrix of the candidate memory unit state; b c Bias vector of candidate memory cell state; C t is the memory cell state at the current time step; f t is the activation value of the forget gate; C t-1 The state of the memory cell at the previous time step;

[0042] The output gate formula is shown in formulas (12)-(13):

[0043] o t =σ(W o ·[h t-1 , x t ]+b o ) (12)

[0044] h t =o t tanh(C t ) (13)

[0045] In formulas (12)-(13): t is the output of the sequence output gate at time t; σ is the RELU activation coefficient; W o is the coefficient weight matrix of the output gate; h t-1 is the hidden state output of the previous sequence LSTM module; x t represents the sample sequence input at time t; b o is the offset term of the output gate; h t is the hidden state output of the LSTM module of the current sequence; C t is the cell state of the sequence at time t;

[0046] σ(z) is the Sigmoid activation function as shown in formulas (14)-(15), and tanh(z) is the tanh activation function as shown in formulas (16)-(17):

[0047]

[0048] σ′(z)=σ(z)[1-σ(z)] (15)

[0049]

[0050] tanh′(z)=1-tanh 2 (z) (17)

[0051] In formulas (14)-(17): σ is the Sigmoid function; z is the input value of the neuron, that is, the result of the linear combination; tanh is the hyperbolic tangent function; e is a natural constant, based on the base of the natural logarithm, approximately 2.71828; σ′(z) is the derivative of the Sigmoid function, used for backpropagation to calculate the gradient; tanh′ is the derivative of the tanh function, used for backpropagation to calculate the gradient;

[0052] Compared with the standard neural network, the neural network with Dropout is equivalent to adding a probability process to the output vector of the previous layer, that is, whether it has been filtered;

[0053] The result evaluation method in step (5) adopts two indicators, NSE as shown in formula (18) and R as shown in formula (19). The evaluation model is evaluated by the NSE and R indicators, and the model effect can be evaluated from the aspects of model simulation effect and goodness of fit. The value range of NSE is (-∞, 1). The closer to 1, the better the prediction effect. The correlation coefficient R measures the proportion of independent variables explained, reflecting the goodness of fit of the regression equation. The goodness of fit of the model is comprehensively evaluated. The closer R is to 1, the greater the proportion of the part that explains the change of the dependent variable value with the independent variable, and the better the fitting effect.

[0054]

[0055] In formula (18): y i and y′ i represent the real observed data and simulated data respectively, Indicates taking the mean of the real observation data, i represents the i-th moment, and n represents the total number of time periods;

[0056]

[0057] In formula (19): y i and y′ i Represents real observed data and simulated data; and represents the mean of the real observation and the mean of the simulated data; i represents the i-th moment; n represents the total number of time periods.

[0058] The SCS model established by the present invention based on the LSTM-SCS data-deficient mountainous area flood process intelligent forecasting method has significant data interpolation effect and high forecast accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0060] Figure 1 This is a flow chart of a method for intelligent flood forecasting in mountainous areas with insufficient data based on LSTM-SCS of the present invention;

[0061] Figure 2 This is the result diagram of intelligent flood forecasting in the Rujigou Basin area of ​​Dawukou District, Ningxia in Example 1. DETAILED DESCRIPTION

[0062] Example 1

[0063] 1. Software Source

[0064] Python 3.8.

[0065] 2. The present invention is a method for intelligent flood forecasting in mountainous areas with insufficient data based on LSTM-SCS

[0066] like Figure 1 As shown, the present invention provides an intelligent forecasting method for flood processes in mountainous areas with insufficient data based on LSTM-SCS, which includes the following steps in sequence:

[0067] (1) Collection and processing of historical flood and rainstorm data and acquisition of underlying surface data in the Rujigou Basin of Dawukou District, Ningxia, at observation sites in mountainous areas with insufficient data

[0068] ①. Collection and processing of historical flood and rainstorm data in the Rujigou River Basin of Dawukou District, Ningxia, at the data-deficient mountain observation station: Collect historical flood and rainstorm data of 13 floods that lasted for at least 24 hours in the Rujigou River Basin of Dawukou District, Ningxia, at the data-deficient mountain observation station. The historical flood and rainstorm data include historical rainfall data, historical flood peak flow and historical peak time; There is no measured rainfall data in the Rujigou River Basin of Dawukou District, Ningxia, at the data-deficient mountain observation station. The measured rainfall data in the surrounding Dawukougou Hydrological and Meteorological Station of Ningxia are used for interpolation to obtain the historical rainfall data of 13 floods that lasted for at least 24 hours in the Rujigou River Basin of Dawukou District, Ningxia, at the data-deficient mountain observation station, arranged in chronological order with time intervals of hours; According to the The historical peak flows of the 13 floods that lasted for at least 24 hours in the Rujigou watershed area of ​​Dawukou District, Ningxia, the data-deficient mountain observation station, were calculated based on flood traces; the historical peak times of the 13 floods that lasted for at least 24 hours in the Rujigou watershed area of ​​Dawukou District, Ningxia, the data-deficient mountain observation station, were determined based on historical records; the DEM data of the Rujigou watershed area of ​​Dawukou District, Ningxia, the data-deficient mountain observation station, were subjected to depression filling and flow direction analysis based on ArcGIS, and the runoff accumulation was analyzed, the river network was extracted and classified, and the watershed basin was segmented to obtain calculation layer analysis data of the watershed basin and river system, and raster river network data was established based on the calculation layer analysis data;

[0069] ② Obtaining underlying surface data for the Rujigou River Basin in Dawukou District, Ningxia, a data-deficient mountain observation site: Based on the engineering facilities, topography, hydrological and meteorological data, vegetation cover, river section and soil type data for the Rujigou River Basin in Dawukou District, Ningxia, a data-deficient mountain observation site, the underlying surface data for the Rujigou River Basin in Dawukou District, Ningxia, a data-deficient mountain observation site, were obtained. The underlying surface data for the Rujigou River Basin in Dawukou District, Ningxia, a data-deficient mountain observation site, are detailed in Table 2.

[0070] Table 2 Data of underlying surface in Rujigou watershed of Dawukou District, Ningxia at observation stations with missing data in mountainous areas

[0071]

[0072] (2) Establishment of the SCS model and completion of flood process lines in the Rujigou Basin of Dawukou District, Ningxia, at observation stations with missing data in mountainous areas

[0073] ①. Establishment of SCS model: a. The SCS model of the Dawukougou watershed in Ningxia, a mountainous area with available data, was established based on the underlying surface data of the Dawukougou watershed in Ningxia, a mountainous area with available data. The underlying surface data of the Dawukougou watershed in Ningxia, a mountainous area with available data, are detailed in Table 3. b. The parameters were adjusted based on the historical flow and rainfall data of 13 floods in the Dawukougou watershed in Ningxia, a mountainous area with available data, which were arranged in chronological order at time intervals of one hour and lasted for at least 24 hours. The parameters were adjusted based on the historical flow and rainfall data of 13 floods in the Dawukougou watershed in Ningxia, a mountainous area with available data, which lasted for at least 24 hours. The historical flow data and historical rainfall data of 13 floods lasting at least 24 hours are shown in Table 4 (taking the 2006 / 07 / 14-15 flood as an example). c. The CN value 87 and the unit line described in step (2) ①b are combined with the underlying surface data of the Rujigou River Basin in Dawukou District, Ningxia, a mountainous observation site with insufficient data, as described in step (1) ② to construct an SCS model of the Rujigou River Basin in Dawukou District, Ningxia, a mountainous observation site with insufficient data; d. The CN value 87 and the unit line described in step (2) ①b are calibrated based on the historical peak flow and historical peak time of 13 floods lasting at least 24 hours in the Rujigou River Basin in Dawukou District, Ningxia, a mountainous observation site with insufficient data, as described in step (1) ①.

[0074] Table 3. Underlying surface data of Dawukougou watershed in Dawukou District, Ningxia, near the observation stations with data in mountainous areas

[0075]

[0076] Table 4 Examples of historical flow and rainfall data for 13 floods lasting at least 24 hours in the Dawukougou Basin of Dawukou District, Ningxia, near observation stations in mountainous areas with available data

[0077] (Take the 2006 / 07 / 14-15 flood as an example)

[0078]

[0079] ②. Completion of flood process lines in the Rujigou Basin area of ​​Dawukou District, Ningxia, at observation stations with missing data in mountainous areas

[0080] According to the SCS model of the Rujigou River Basin in Dawukou District, Ningxia, the data-deficient mountain observation site described in step (2)① and the historical rainfall data of the Rujigou River Basin in Dawukou District, Ningxia, the data-deficient mountain observation site described in step (1)①, the historical peak flows and flood process lines of the historical peak times of the 13 floods that lasted for at least 24 hours in the Rujigou River Basin in Dawukou District, Ningxia, the data-deficient mountain observation site described in step (1)① are completed to obtain the historical flow data of the 13 floods that lasted for at least 24 hours in the Rujigou River Basin in Dawukou District, Ningxia, the data-deficient mountain observation site, arranged in chronological order with time intervals of hours; Table 5 is a comparison table of the changes in the number of historical flow data before and after the completion of the flood process line data of the 13 floods that lasted for at least 24 hours in the Rujigou River Basin in Dawukou District, Ningxia, the data-deficient mountain observation site;

[0081] Table 5 Comparison of the changes in the number of historical flow data before and after the completion of the 13 flood process line data lasting at least 24 hours in the Rujigou Basin of Dawukou District, Ningxia, at the mountainous observation station with missing data

[0082]

[0083] (3) Preprocessing of input data of LSTM neural network prediction model

[0084] Step (1) ① The 13 flood historical rainfall data of at least 24 hours in the Rujigou River Basin of Dawukou District, Ningxia, which are arranged in chronological order at time intervals of hours at the data-deficient mountain observation station, are arranged in chronological order. The 7 rainfall data P1-P7 of each flood rainfall data are arranged in chronological order to form a single rainfall data group for each flood. The P7 moment of each flood single rainfall data group has a certain flow data Q of each flood flow data in the historical flow data corresponding to step (2) ②; the certain flow data Q corresponding to the P7 moment of each flood single rainfall data group and P1-P7 of each flood single rainfall data group are arranged in sequence to form a single rainfall flow data group for each flood; at least 18 of the single rainfall data groups for each flood or the single rainfall flow data groups for each flood are arranged in chronological order. The data groups are arranged in sequence according to the time sequence of P1 to form each flood rainfall data group or each flood rainfall flow data group; the single rainfall data groups of each flood in the order of each flood rainfall flow data group constitute each flood time series data set, and each data in the each flood time series data set is scaled to (0, 1) using max-min normalization as input data for the LSTM neural network prediction model of each flood; Table 6 shows the number of LSTM input rainfall flow data groups for 13 floods that lasted at least 24 hours in the Rujigou River Basin of Dawukou District, Ningxia, a data-deficient mountainous observation site; Table 7 shows an example of 13 flood rainfall flow group data that lasted at least 24 hours in the Rujigou River Basin of Dawukou District, Ningxia, a data-deficient mountainous observation site (taking the 2006 / 07 / 14-15 flood as an example);

[0085] Table 6 Number of LSTM input rainfall flow data sets for 13 floods lasting at least 24 hours in the Rujigou watershed area of ​​Dawukou District, Ningxia at observation sites in mountainous areas with missing data

[0086]

[0087] Table 7 Examples of rainfall flow group data for 13 floods lasting at least 24 hours in the Rujigou Basin of Dawukou District, Ningxia at observation sites in mountainous areas with missing data

[0088] (Take the 2006 / 07 / 14-15 flood as an example)

[0089]

[0090] (4) Establishment of LSTM neural network prediction model

[0091] The parameters of the LSTM neural network model were preliminarily formulated, including the number of hidden layers being 2, the number of nodes in the first hidden layer and the second hidden layer being 60, and the number of epoch iterations being 1000. In the case of limited training data, deep neural networks are prone to over-learning and overfitting. If overfitting occurs in deep neural networks, it will have very serious consequences. Dropout can prevent overfitting by randomly discarding units during neural network training. The Dropout of LSTM is set to 0.2. Ten floods out of the 12 floods, which are the sum of the number of floods in each training set and validation set after the preprocessing of the input data of the LSTM neural network prediction model in step (3), are used as training sets, and two floods are used as validation sets. The training set is input into the LSTM model for iterative training, and the training results are verified using the validation set until the prediction error meets the preset standard, and a prediction error sequence is output.

[0092] (5) Evaluation of prediction results

[0093] Among the 13 floods, in addition to the training set and validation set in step (4), another flood on July 23, 2018 was forecasted. The NSE values ​​and R values ​​of the flood prediction results of the LSTM-SCS neural network, BPNN-SCS neural network and GRU-SCS neural network at the Rujigou observation station in Dawukou District, Ningxia, which lacks data, are shown in Table 8 and Figure 2 .

[0094] Table 8 NSE and R values ​​of flood prediction results of LSTM-SCS neural network, BPNN-SCS neural network and GRU-SCS neural network in Rujigou Basin of Dawukou District, Ningxia at observation stations in mountainous areas with missing data

[0095] Neural network name LSTM-SCS neural network BPNN-SCS neural network GRU-SCS neural network NSE value 0.88 0.81 0.77 R-value 0.78 0.73 0.62

[0096] From Table 8 and Figure 2 It can be seen that the NSE value and R value of the flood prediction results of the LSTM-SCS neural network are both the highest, indicating that the intelligent forecasting method for flood processes in mountainous areas with insufficient data based on the LSTM-SCS neural network model of the present invention has higher prediction accuracy and is more suitable for forecasting flood processes in mountainous areas with insufficient data.

[0097] The embodiment illustrates that the method for intelligently predicting flood processes in mountainous areas with insufficient data using an LSTM-SCS neural network model of the present invention establishes an SCS model of the Dawukougou watershed area in the neighboring mountainous areas with sufficient data through the underlying surface data of the Dawukougou watershed area in the neighboring mountainous areas with sufficient data, and adjusts the parameters based on the historical flow data and historical rainfall data of 13 floods that lasted for at least 24 hours in the Dawukougou watershed area in the neighboring mountainous areas with sufficient data at time intervals of hours and arranged in chronological order to obtain a suitable CN value 87 and a unit line; the CN value 87 and the unit line are compared with the data-deficient mountainous area observation line obtained by the method for intelligently predicting flood processes in mountainous areas with insufficient data using an LSTM-SCS neural network model of the present invention. The underlying surface data of the Rujigou River Basin in Dawukou District, Ningxia, the observation station with missing data, was combined to construct the SCS model of the Rujigou River Basin in Dawukou District, Ningxia, the observation station with missing data, and the CN value 87 and the unit line were calibrated according to the historical peak flow and historical peak time of 13 floods that lasted for at least 24 hours; based on the SCS model of the Rujigou River Basin in Dawukou District, Ningxia, the observation station with missing data, the flood process lines of the historical peak flow and historical peak time of 13 floods that lasted for at least 24 hours were completed, and the Rujigou River Basin in Dawukou District, Ningxia, the observation station with missing data, arranged in chronological order with time intervals of hours, were obtained. The historical flow data of 13 floods lasting at least 24 hours, the minimum number of completed historical flow data in Example 1 is 23, and the maximum number of completed historical flow data reaches 47, which meets the requirement of the number of input data for the LSTM neural network prediction model; in the preprocessing step of the input data of the LSTM neural network prediction model, the single rainfall flow data group of each flood session adopts a certain flow data Q of each flood flow data in the historical flow data corresponding to the P7 moment of each flood single rainfall data group and the seven rainfall data of each flood single rainfall data group P1-P7 arranged in sequence with time intervals of hours, and each flood rainfall flow data group is composed of The sorted single rainfall flow data sets for each flood event constituted a flood time series dataset. Max-min normalization was used to scale each data point in the flood time series dataset to (0, 1) as input data for the LSTM neural network prediction model. Ten of the 12 floods, which included the sum of the number of floods in the training and validation sets after preprocessing the input data for the LSTM neural network prediction model, were used as the training set, and two floods were used as the validation set. A single flood from the 2018 / 07 / 23 flood, not included in the training and validation sets, was forecasted. The flood prediction results of the LSTM-SCS neural network showed NSE values ​​and R values ​​of 0.88 and 0.78, both higher than the NSE and R values ​​of the flood prediction results of the BPNN-SCS neural network and the GRU-SCS neural network, indicating that the SCS model established by the LSTM-SCS method for intelligent flood forecasting in mountainous areas with insufficient data has a significant effect on interpolation of insufficient data and has the characteristics of high forecast accuracy.

Claims

1. A method for intelligent flood forecasting in mountainous areas with insufficient data based on LSTM-SCS, comprising the following steps: (1) Collection and processing of historical flood and rainstorm data in the basin area of ​​mountainous observation stations with insufficient data and acquisition of underlying surface data ①. Collection and processing of historical flood and rainstorm data in the watershed of the observation station in the mountainous area with lack of data: collect historical flood and rainstorm data of at least 12 floods that lasted for at least 24 hours in the watershed of the observation station in the mountainous area with lack of data, and the historical flood and rainstorm data include historical rainfall data, historical flood peak flow and historical peak time; if there is no rainfall measured data in the watershed of the observation station in the mountainous area with lack of data, the rainfall measured data in the watershed of the surrounding observation station is used for interpolation to obtain historical rainfall data of at least 12 floods that lasted for at least 24 hours in the watershed of the observation station in the mountainous area with lack of data at time intervals of hours and arranged in chronological order; according to the data of the observation station in the mountainous area with lack of data The historical peak flows of at least 12 floods lasting at least 24 hours in the watershed of the data-deficient mountain observation station are calculated based on flood traces in the watershed of the data-deficient mountain observation station; the historical peak times of at least 12 floods lasting at least 24 hours in the watershed of the data-deficient mountain observation station are determined based on historical records; the DEM data of the watershed of the data-deficient mountain observation station are subjected to depression filling and flow direction analysis based on ArcGIS, and the accumulated runoff is analyzed, the river network is extracted and classified, and the watershed basin is segmented to obtain calculation layer analysis data of the watershed basin and river system, and raster river network data is established based on the calculation layer analysis data; ② Obtaining underlying surface data for watersheds of observation stations in mountainous areas with insufficient data: Based on the data on engineering facilities, topography, hydrology and meteorology, vegetation cover, river sections and soil types in the watersheds of observation stations in mountainous areas with insufficient data, the underlying surface data for the watersheds of observation stations in mountainous areas with insufficient data are obtained; (2) Establishment of SCS model and completion of flood process lines in the basin area of ​​mountainous area with missing data ①. Establishment of SCS model: a. Establish an SCS model for the watershed area of ​​the adjacent mountainous observation station with data based on the underlying surface data of the watershed area of ​​the adjacent mountainous observation station with data; b. Rate and adjust the parameters to obtain a suitable CN value and unit line based on the historical flow data and historical rainfall data of at least 12 floods with time intervals of hours and arranged in chronological order for at least 24 hours in the watershed area of ​​the adjacent mountainous observation station with data; c. Combine the CN value and unit line described in step (2) ①b with the underlying surface data of the watershed area of ​​the mountainous observation station with insufficient data in step (1) ② to construct an SCS model for the watershed area of ​​the mountainous observation station with insufficient data; d. Rate the CN value and unit line described in step (2) ①b based on the historical flood peak flow and historical peak time of the watershed area of ​​the mountainous observation station with insufficient data in step (1) ①; ②. Completion of flood process lines in the watershed area of ​​the mountainous observation station with missing data: Based on the SCS model of the watershed area of ​​the mountainous observation station with missing data in step (2)① and the historical rainfall data of the watershed area of ​​the mountainous observation station with missing data in step (1)①, the flood process lines of the historical peak flow and historical peak time of at least 12 floods lasting at least 24 hours in the watershed area of ​​the mountainous observation station with missing data in step (1)① are completed to obtain the historical flow data of at least 12 floods lasting at least 24 hours in the watershed area of ​​the mountainous observation station with missing data, which are arranged in chronological order with time intervals of hours; (3) Preprocessing of input data of LSTM neural network prediction model Step (1) ① The seven rainfall data P1-P7 of each flood rainfall data in the historical rainfall data of the watershed area of ​​the data-deficient mountain observation station are arranged in chronological order to form a single rainfall data group for each flood, and the P7 moment of each single rainfall data group for each flood has a certain flow data Q of each flood flow data in the historical flow data corresponding to step (2) ②; the certain flow data Q corresponding to the P7 moment of each single rainfall data group for each flood and P1-P7 in each single rainfall data group for each flood are arranged in sequence to form a single rainfall data group for each flood. Flow data group; at least 18 of the single rainfall data groups of each flood or the single rainfall flow data groups of each flood are arranged in sequence according to the time of P1 in the data group to form each flood rainfall data group or each flood rainfall flow data group; the single rainfall data groups of each flood sorted in the each flood rainfall flow data group constitute each flood time series data set, and each data in the each flood time series data set is scaled to (0, 1) using max-min normalization as input data for each flood LSTM neural network prediction model; (4) Establishment of LSTM neural network prediction model The parameters of the LSTM neural network model are preliminarily formulated, including the number of hidden layers, the number of hidden layer nodes, and the number of epoch iterations. The forward propagation calculation formula of the Dropout neural network is added to the standard LSTM neural network as shown in formulas (1)-(4). 80% of the total number of floods in the training set and validation set after the preprocessing of the input data of the LSTM neural network prediction model in step (3) is used as the training set, and 20% as the validation set. The training set is input into the LSTM model for iterative training, and the training results are verified using the validation set until the prediction error meets the preset standard, and the prediction error sequence is output. M(l)~Bernoulli(p)(1)y ～(l) =r(l)*x(l)(2) In formulas (1)-(4), M is a random mask matrix; l represents the layer index of the neural network, and the lth layer represents the current layer; for a neural network, the layers are arranged in order, from the input layer to the output layer; p is the retention rate; y ～(l) represents the output of layer l after the dropout operation; r(l) represents the dropout mask of layer l, expressed as a vector of 0 or 1; x(l) represents the input of layer l; represents the linear combination input of the i-th neuron in the l+1th layer; is the weight vector connecting the i-th neuron in the l-th layer and the l+1-th layer; represents the bias of the i-th neuron in the l-th layer; represents the activation output of the i-th neuron in the l+1th layer; f is the activation function of the gate; (5) Evaluation of prediction results The flood process prediction results of the LSTM-SCS neural network are compared and analyzed with those of the BPNN-SCS neural network and the GRU-SCS neural network using the NSE and R indicators. The higher the values ​​of the NSE and R indicators, the higher the accuracy of the prediction results.

Images