CN114386677A - Flood forecasting method based on novel universal input/output structure and long-and-short time memory network - Google Patents

Abstract

A flood forecasting method based on a novel universal input and output structure and a long-time and short-time memory network is disclosed. Firstly, analyzing the confluence characteristic according to the historical flood data of the research basin, and calculating the average confluence time of the research basin; and secondly, determining the number of output time segments of the LSTM flood forecasting model according to the average convergence time of the drainage basin, and giving the number of hidden layer layers and the number of hidden layer neuron nodes of the LSTM flood forecasting model. And then, designing an input/output structure of the novel universal LSTM flood forecasting model, inputting a training set and a verification set sample training model, and obtaining the flood forecasting models with different structures and parameters. And finally, comparing and analyzing the performances of the LSTM flood forecasting models under different input lengths in a training set and a verification set, determining the final superior LSTM flood forecasting model, and evaluating and analyzing the forecasting effect of the LSTM flood forecasting model in a testing set. The method has strong universality, and the established LSTM flood forecasting model can obtain a good forecasting effect and provide a new technical support for the flood disaster defense work of the drainage basin.

Description

A Flood Forecasting Method Based on Novel Universal Input-Output Structure and Long Short-Term Memory Network

technical field

The invention belongs to the technical field of flood forecasting in river basins, and relates to a flood forecasting method based on a novel general input-output structure and a long-short-term memory network.

Background technique

Long short-term memory network (LSTM) is one of the most commonly used models in the cross-application research of machine learning technology and watershed flood forecasting. LSTM flood forecasting model takes samples as input and output data, samples are the combination of input feature variables and corresponding output target values, and the generation of samples only depends on the input and output structure of the model. Determining the input and output structure of the model is a key step in building a machine learning flood forecasting model, and it is more directly related to the reliability and rationality of the LSTM flood forecasting model. In the past research and application of flood forecasting based on machine learning technology, precipitation and flow are the two most commonly used input feature variables in machine learning flood forecasting models. However, the actual application results show that when the current measured rainfall and flow are both used as the model input characteristic factors, due to the significant correlation between flows at different time steps, the output flow is easily driven by the previous flow in the input characteristic variables. It is difficult to identify the role of precipitation and flow in the early stage, which leads to the phenomenon of time lag between peak occurrence time or output of two flood peaks in the process of simulating or forecasting flow, and the longer the forecast period, the more obvious the delay of flood peak. Therefore, the flood forecasting model established with the structure of rainfall and pre-flow flow as input and forecast flow target as output may not be consistent with the mechanism of runoff formation in the basin. It is necessary to redesign the input and output structure of the LSTM flood forecasting model to enhance the machine Learn about the interpretability of flood forecasting models.

To this end, the present invention is based on improving the hydrological interpretability of the machine learning flood forecasting model, and proposes a flood forecasting method based on a novel general input-output structure and a long-short-term memory network. Combined with the theory and method of watershed hydrological simulation and the response mechanism of rainstorm and flood in the watershed, rainfall is selected as the only input characteristic variable of the LSTM flood forecasting model, and the long-sequence rainfall information of each rainfall station in the watershed is used as the model input variable (the number of characteristic factors is the number of characteristic factors in the watershed). The number of rainfall stations), relying on the recursive connection of LSTM and the state of special control gates and cell units, give full play to the potential long-term learning and memory ability of LSTM, through repeated learning of long-sequence input data, the spatial and temporal distribution of rainfall in the basin, the rainstorm flood in the basin The information of hydrological elements such as response time and previous soil water storage status is integrated into the input and output structure design of the model. On this basis, a flood forecasting model based on long and short-term memory network is constructed, which enriches the flood forecasting model in areas with data and improves the flood forecasting in the basin. Forecast accuracy.

SUMMARY OF THE INVENTION

Aiming at the problems existing in the prior art, the present invention provides a flood forecasting method based on a novel general-purpose input-output structure and a long-short-term memory network.

In order to achieve the above object, the technical scheme adopted in the present invention is as follows:

A flood forecasting method based on a novel general input-output structure and long-short-term memory network, including the following steps:

The first step is to collect and organize the historical flood data of the basin.

Collect and organize the flood data of the research watershed events, and divide all flood events into training set, validation set, and test set. Among them, the training set floods are used to optimize the internal weight matrix and bias vector parameters of the LSTM flood forecasting model; the validation set floods are used to assist in determining the external settings of the model such as hyperparameters and loss functions; the test set floods are used to test the training The extrapolation (prediction) ability of the post model.

The second step is to calculate the average confluence time of the watershed.

The average confluence time of a certain research basin is determined, and its duration is equal to the forecast period of flood forecasting in the basin, which comprehensively reflects the regulation and storage effect of the basin on the convergence process of rainfall and water flow. According to the historical flood process data collected and organized in the watershed, the time difference between the main rainfall and the corresponding flood peak flow in the flood events is counted, that is, the flood peak delay time. Calculate the average peak delay time of all floods, which is the average confluence time of the basin.

In the third step, the number of hidden layer layers and the number of hidden layer neurons in the flood prediction model of the Long Short Term Memory (LSTM) network are given.

The fourth step is to design the input and output structure of the new general LSTM flood forecasting model.

The LSTM cell unit state is similar to the watershed soil water storage state. The interaction between the three control gates (forgetting gate, input gate and output gate) and the cell unit state can be regarded as the consumption, increase and release of the watershed soil water storage state. According to the traditional watershed hydrological simulation theory and method, rainfall is selected as the only input factor of the LSTM flood forecast model. The input of the LSTM flood forecast model is the long-sequence rainfall information of each rainfall station in the watershed, and the input length is n+l time periods. Among them, n represents the input length of the previous rainfall, and the previous rainfall can be regarded as reflecting the influence of information such as the soil water storage state and short-term rainfall in the basin on the output flow value of the model. Multiple n can be selected and the better one can be determined according to the subsequent model performance. Model; the output of the LSTM flood forecast model is a sequence of flow values equal to the confluence time of the study watershed (the calculation result of the second step), that is, the output length is l time periods, where l is equal to the watershed confluence time and l≤n.

The fifth step is to generate training, validation and test sample sets.

The sample length is determined according to the input and output structure (values of n and l) of the new general LSTM flood forecasting model designed in the fourth step. The length of the input rainfall sequence of each sample is equal to n+l periods, and the length of the output target flow sequence is equal to l period. According to the training set, validation set and test set divided in the first step, the corresponding training, validation and test sample sets are generated for each flood according to the method of sliding interception by time period, and each flood generates multiple samples, each sample From the input rainfall sequence P=[P _t-n+1 ,P _t-n+2 ,…,P _t ,…,P _t+1 ] and the output target flow sequence Q=[Q _t+1 ,Q _t+2 ,…,Q _t ,…,Q _t+l ] constitute the input and output data pairs. Among them, n is the number of periods of the input precipitation sequence of the LSTM flood forecast model, l is the number of periods of the model output flow, and the input rainfall Pt of the _t -th period includes the measured rainfall value of each rainfall station in the basin.

The sixth step is model construction and training.

The LSTM flood forecasting model uses the back-propagation over time (BPTT) algorithm for supervised learning training. The model construction and training are based on the open source Keras and TensorFlow implementations. The Keras framework running on the TensorFlow platform integrates a relatively mature machine learning algorithm package, and can directly call the corresponding algorithm to complete the construction and training of the LSTM flood forecasting model.

6.1) LSTM flood forecasting model construction: According to the number of hidden layers given in the third step, the number of hidden layer neurons and the model input and output structure designed in the fourth step, call Keras's layer package (layers) to define the input layer of LSTM , hidden layer and output layer to build an LSTM flood forecasting model;

6.2) LSTM flood forecasting model training: input the training and verification sample sets generated in the fifth step, set the hyperparameters, activation functions, loss functions, optimization algorithms, etc. involved in the model training process, and run the LSTM flood based on the Keras framework on the TensorFlow platform Forecast model, get the trained LSTM flood forecast model.

The seventh step is to determine the optimal model, and extract and analyze the flood simulation results of the watershed.

Compare and analyze the performance of the LSTM flood forecasting model under different input lengths, determine the final optimal LSTM flood forecasting model, further extract the results of flood simulation and forecast flow process in the basin, and analyze the forecasting performance of the LSTM flood forecasting model.

The invention introduces a long-short-term memory network as a modeling tool for flood forecasting in the basin, innovatively designs an input-output structure with long-sequence rainfall as input and multi-step flow as output, and learns and remembers long-sequence rainfall implicits in cell unit state. It provides theoretical and methodological support to provide theoretical and methodological support for early-stage soil water storage status information in the basin, and control gates to adjust the dynamic changes of the basin soil water storage status, consolidate the interpretability foundation of the machine learning flood forecasting model, and build a flood forecasting model based on LSTM on this basis. In order to improve the accuracy of flood forecasting in the basin.

The above-mentioned flood forecasting method based on a new general input-output structure and long-short-term memory network is applied to flood forecasting in the basin.

The beneficial effects of the present invention are:

Traditional machine learning flood forecasting models mostly use rainfall and flow as input variables, and the model cannot really play the role of rainfall on forecast flow in the input variables, resulting in poor hydrological interpretation of machine learning flood forecasting models. The invention deeply analyzes the basic principle and computing mechanism of the special control gate and cell unit state inside the long-short-term memory network, and designs a new general input and output structure of the LSTM flood forecast model with long-sequence rainfall input and multi-step flow output. It can be applied to flood forecasting and modeling of watersheds at different spatial scales, fully utilizes the long-term learning and memory ability of LSTM and the information forgetting, storing and releasing mechanism of internal cell units, and realizes the instantiated application of the LSTM flood forecasting model in mountainous watersheds. The simulation and forecast accuracy of floods in the basin has been improved, and new technical support has been provided for the forecast and early warning of flood disasters in the basin.

Description of drawings

Fig. 1 is the Anhe watershed map that the example application of the present invention adopts;

Fig. 2 is the hydrological meaning schematic diagram of the internal structure of LSTM of the present invention;

Fig. 3 is the input and output structure schematic diagram of the LSTM flood forecasting model designed by the present invention;

Figure 4a is a schematic diagram of a sample generated at the t-th moment of flood in the present invention;

Figure 4b is a schematic diagram of a sample generated at time t+1 of a flood in the present invention;

5 is a schematic diagram of the hierarchical structure and matrix feature dimension of the LSTM flood forecasting model of the present invention;

6 is a schematic diagram of the training process of the LSTM flood forecasting model of the present invention;

Fig. 7 (a) is the NSE mean value diagram of different input rainfall length training set sessions of the LSTM flood forecasting model of the present invention;

Fig. 7(b) is the MAE mean value diagram of the training set sessions of different input rainfall lengths of the LSTM flood forecasting model of the present invention;

Fig. 7 (c) is the RMSE mean value diagram of different input rainfall length training set sessions of the LSTM flood forecasting model of the present invention;

Fig. 7 (d) is the NSE mean value diagram of different input rainfall length test sets of the LSTM flood forecasting model of the present invention;

Fig. 7(e) is the MAE mean value diagram of different input rainfall length test sets of the LSTM flood forecasting model of the present invention;

Fig. 7(f) is the RMSE mean value diagram of different input rainfall length test sets of the LSTM flood forecasting model of the present invention;

8 is a schematic diagram of the input and output structure of the preferred LSTM flood forecasting model in the application of the example of the present invention;

Figure 9(a) is a comparison diagram of the simulation flow process of the training set session 19840501 of the LSTM flood forecasting model of the present invention;

Figure 9(b) is a comparison diagram of the simulation flow process of training set 19900730 of the LSTM flood forecasting model of the present invention;

Figure 9 (c) is a comparison diagram of the simulation flow process of the training set session 19940613 of the LSTM flood forecasting model of the present invention;

Figure 9(d) is a comparison diagram of the simulation flow process of the training set 19970620 of the LSTM flood forecasting model of the present invention;

Figure 9(e) is a comparison diagram of the simulation flow process of the training set session 20020805 of the LSTM flood forecasting model of the present invention;

Figure 9(f) is a comparison diagram of the simulated flow process of the test set field 20060505 of the LSTM flood forecasting model of the present invention;

Fig. 9 (g) is the test set field 20060725 simulation flow process comparison diagram of the LSTM flood forecasting model of the present invention;

Figure 9 (h) is a comparison diagram of the simulated flow process of the test set field 20080524 of the LSTM flood forecasting model of the present invention;

Fig. 9 (i) is the test set field 20080608 simulation flow process comparison diagram of the LSTM flood forecasting model of the present invention;

Figure 9(j) is a comparison diagram of the simulated flow process of the test set 20120621 of the LSTM flood forecasting model of the present invention.

Detailed ways

The present invention will be further described below in conjunction with specific embodiments.

The invention proposes a flood forecasting method based on a novel general input-output structure and a long-short-term memory network. The flood results of the training set and test set in the application of the watershed example represent the simulation and prediction performance of the LSTM flood forecasting model, respectively. The present invention will be further described below through implementation cases and in conjunction with the accompanying drawings.

The Anhe River Basin is located in Shangyou County, Ganzhou City, Jiangxi Province, at 114°-114°40'E, 25°42'-26°01'N, with a drainage area of 251km ² . The condition of hydrological observation data in Anhe Basin is good, which can provide better data support for the construction of machine learning flood forecasting model. In addition, the vegetation in the basin is well-developed and belongs to the subtropical hilly and mountainous humid monsoon climate zone, with abundant rainfall, with an average annual rainfall of 1497 mm and an average temperature of 18.8 °C. The number of rainfall stations in the Anhe watershed is 8, and the spatial distribution is relatively uniform. The hill basin is selected as a research example for flood forecasting, and flood forecasting based on long and short-term memory network is realized. The main steps are as follows:

The first step is to collect and organize the historical flood data of the basin.

The measured hydrological data (8 rainfall stations, 1 hydrological station) in the Anhe River Basin from 1984 to 2012 were collected, and a few floods with unbalanced water volume were eliminated by comparing the measured rainfall and flow data sequences over the years to ensure Reliability of hydrological data. At the same time, there are problems in the collected rainfall and flow data, such as errors in the logical relationship between the start and end times, and duplication of observation records, which need to be manually identified and corrected. In addition, the rainfall data is time-by-period data, and the start and end time intervals are not uniform, while the flow data is instantaneous data. It is necessary to carry out pre-processing such as rationality inspection and time-period interpolation of the rainfall and flow data, and organize the rainfall and flow data into hourly accumulation. Rainfall, mean flow time series. On this basis, according to the sequence of flood occurrences, the first 50% of all floods are selected as the training set, the middle 20% as the validation set, and the ratio of floods in training, validation and test sets is 5:2:3.

The second step is to calculate the average confluence time in the Anhe watershed.

According to the precipitation flow process data of flood events in Anhe watershed, the lag time of flood peak flow in most flood events in Anhe watershed is 4-6 hours, and the average lag time is 5.21 hours. Accordingly, the average confluence time in the Anhe River Basin is determined to be 6 hours, that is, the forecast period of flood forecasting is 6 hours.

In the third step, the hidden layer of the given LSTM flood forecasting model is a single layer, and the number of neurons in the hidden layer is 5.

The fourth step is to design the input and output structure of the new general LSTM flood forecasting model.

The LSTM cell unit state is similar to the watershed soil water storage state. The interaction between the three control gates (forgetting gate, input gate and output gate) and the cell unit state can be regarded as the consumption, increase and release of the watershed soil water storage state. Figure 2 shows the input and output calculation process of the LSTM model in the t-th (current) period and the hydrological meaning of the model input variables. The input and output structure of designing the LSTM flood forecasting model is shown in Figure 3. Among them, the output of the model is the flow value Q=[Q _t+1 ,Q _t+2 ,…,Q _t+6 ] in different time periods during the forecast period of flood forecasting in the Anhe Basin, and the input is the long-sequence actual measurement of each rainfall station in the basin Rainfall (pre-rain + short-term rainfall). The previous rainfall sequence reflects the impact of changes in the basin state, such as soil water storage capacity, on the output flow. The specific selection period of the previous rainfall should be determined by a trial algorithm based on the characteristics of the basin. The LSTM flood forecast model can be constructed by setting the input scheme of the series of early rainfall (24h, 48h, 120h, 240h, 480h, 720h, 960h and 1440h), and the simulation and prediction performance of the model can be compared to determine the optimal order of the previous rainfall input.

The fifth step is to generate training, validation and test sample sets.

According to the training, verification and test set sessions of the research watershed divided in the first step, and according to the model input and output structure corresponding to the different previous rainfall input schemes in the fourth step, each flood is generated according to the method of sliding interception period by period. Taking the model input rainfall of 16h and output target flow of 6h as an example, Figure 4 is an example diagram of generating machine learning model samples from the adjacent time period t to t+1 of the flood process. The number of samples in the Anhe watershed training set, validation set and test set is shown in Table 1.

Table 1 Number of samples in Anhe watershed training set, validation set and test set

The sixth step is to build and train the LSTM flood forecasting model.

The LSTM model uses the back-propagation over time (BPTT) algorithm for supervised learning training. The model construction and training are based on the open source Keras and TensorFlow implementations. According to the model structure settings determined in the third and fourth steps, taking the previous rainfall input of 480h as an example, Figure 5 is a visualization of the hierarchical structure of the LSTM flood forecasting model based on TensorFlow and Keras and the feature dimensions of input and output variables. Among them, the model input is the long sequence (486h) measured rainfall of 8 rainfall stations in the Anhe Basin; the model output is the flow value corresponding to the last 6 periods of the long sequence; the function of the TimeDistributed layer is to connect the hidden states h _t at each moment. To the output layer (Dense) where the number of neuron nodes is 1, the flow value is output at each moment; the Lambda layer is a custom layer, which realizes the slice processing of the output flow value sequence data, so as to extract the last 6 time periods of each sample The output flow value is used as the final output sequence of the model.

Figure 6 is a schematic diagram of the training process of the LSTM flood forecasting model under the TensorFlow deep learning framework. Among them, the hyperparameters epoch=300, batchsize=64, the loss function is set to mean square error (MSE), the activation function of the output layer adopts the linear ReLU function, the optimization algorithm adopts the Adam algorithm, and the learning rate is 0.0006. The MSE formula is shown in formula (1).

In the formula: y _k,obs and y _k,out are the measured flow value and forecast flow value of a flood in the kth period, respectively, in m ³ /s; l is the output length of the simulated flow. The ReLU activation function has the advantages of fast convergence and simple calculation, and can ensure that the output flow values of the LSTM flood forecast model are all non-negative values. The calculation expression is shown in Equation (2).

ReLU(x)=max(x,0)(2)

The seventh step is to determine the optimal model, and extract and analyze the flood simulation results of the watershed.

Construct the corresponding LSTM flood forecasting models under different previous rainfall input schemes, and compare and analyze the flood simulation and prediction of LSTM models under different previous rainfall input schemes (24h, 48h, 120h, 240h, 480h, 720h, 960h and 1440h) in the Anhe watershed The Nash efficiency coefficient (NSE), mean absolute error (MAE), and root mean square error (RMSE) calculations for the flow process are shown in Figure 7. The calculation expressions of NSE, MAE and RMSE are as follows:

In the formula: y _k,obs and y _k,out are the measured flow value and forecast flow value of a flood in the kth period, respectively, in m ³ /s; l is the output length of the simulated flow.

According to Figure 7, when the current rainfall input is ≥480h, the mean values of NSE, MAE, and RMSE of the training set and validation set of flood simulation and predicted flow process are significantly better than those of the previous rainfall input ≤240h, and the length of the previous rainfall is further extended ( 720h, 960h and 1440h) failed to improve the results of the model. The input and output structure of the final optimal LSTM flood forecasting model in the Anhe watershed is thus determined as shown in Figure 8.

The finally established LSTM flood forecasting model was applied to flood forecasting in the Anhe watershed, and the flood calculation results of the training set and test set were extracted and analyzed. In addition to NSE, MAE and RMSE, the peak pass rate (QRP) and peak time pass rate (QRT) indicators are used to evaluate the model's quality of forecast results for flood peak flow. The calculation of QRP and QRT is shown in formula (6) and formula (7):

In the formula: NP represents the qualified quantity of flood peak flow of the event, N _T represents the qualified number of flood peak current time of the event, and N represents the total number of floods of the event _. Taking the flood simulation and prediction results corresponding to the sixth period of model output as an example, Table 2 shows the flood simulation results of the LSTM flood forecast model, and takes the traditional Xin'anjiang conceptual hydrological model as a comparison benchmark. It can be seen from Table 2 that the peak simulation and prediction results of the LSTM model are significantly better than those of the XAJ model. Compared with the XAJ model, the peak pass rate QRP of the floods in the training set of the LSTM model was increased from 77% to 82%, and the peak pass rate QRP of the floods in the test set was increased from 72% to 80%; the average NSE of the floods in the training set was increased from 0.825 When it is increased to 0.871, the mean NSE of the floods in the test set is increased from 0.815 to 0.821; the evaluation indicators MAE and RMSE of the floods in the training set and test set of the LSTM model are also lower. The analysis of the above results shows that the constructed LSTM flood forecast model has successfully established the complex nonlinear relationship between rainfall and outlet flow in the basin, and the model has high flood forecast accuracy in the Anhe basin.

Table 2 Statistical results of flood simulations in the LSTM flood forecasting model in Anhe Basin

Taking the flood simulation and predicted flow process corresponding to the sixth period of model output as an example, five training sets and test sets of floods were selected respectively, and the flood simulation, prediction and measured flow processes of the two flood forecast models, LSTM and XAJ, were drawn. The comparison chart is shown in Figure 9. It can be seen from Figure 9 that compared with the results of the XAJ model, the simulated and predicted flow processes of the LSTM flood forecast model are more consistent with the measured flow, and the rise and fall stages of floods are basically consistent with the results of the XAJ model. The constructed LSTM flood forecast model fully learned the conversion mechanism between rainfall and runoff in the basin, and established the nonlinear mapping relationship between rainfall and outlet flow.

The above results show that the flood forecasting method based on the novel general-purpose input-output structure and long-short-term memory network proposed by the present invention can successfully guide the LSTM flood forecasting model to mine and learn the complex nonlinear mapping relationship between rainfall and outlet flow in the basin , the fitting ability of the model to the relationship between rainfall and flow, and the forecasting ability of the flood peak value of the flood events are better. In addition, the designed new general input and output structure has strong practicability in flood forecasting and modeling of watersheds at different spatial scales. The machine learning flood forecasting model established based on this input and output structure can give full play to the influence of rainfall and water storage status in the watershed. Floods play a contributing role in rising and retreating, and their hydrological interpretability is based on a strong foundation.

The above-mentioned embodiments only represent the embodiments of the present invention, but should not be construed as a limitation on the scope of the present invention. It should be pointed out that for those skilled in the art, without departing from the concept of the present invention, Several modifications and improvements can also be made, which all belong to the protection scope of the present invention.

Claims (1)

1. a flood forecasting method based on a novel general input-output structure and a long-short-term memory network, characterized in that the hydrological implication of the flood forecasting method is significant, and the model input-output structure can be applied to the flood forecasting of river basins of different spatial scales, Include the following steps: The first step is to collect and organize the historical flood data of the basin; Collect and organize the flood data of the research watershed, and divide all the floods into training set, validation set, and test set. The training set flood is used to optimize the internal weight matrix and bias vector parameters of the LSTM flood forecasting model; The floods are used to assist in determining the external settings of the model such as hyperparameters and loss functions; the floods in the test set are used to test the extrapolation ability of the model after training; The second step is to calculate the average confluence time of the watershed; The average confluence time of a certain research basin is determined, and its duration is equal to the forecast period of flood forecasting in the basin, which comprehensively reflects the regulation and storage effect of the basin on the convergence process of rainfall and water flow; The time difference between the main rainfall and the corresponding flood peak flow in a flood event is the peak delay time; the average value of the flood peak delay time of all flood events is calculated, which is the average confluence time of the basin; The third step is to give the number of hidden layers and the number of hidden layer neurons in the flood prediction model of the long-short-term memory LSTM network; The fourth step is to design the input and output structure of the new general LSTM flood forecasting model; The LSTM cell unit state is similar to the watershed soil water storage state. The interaction between the three control gates and the cell unit state can be regarded as the consumption, increase and release of the watershed soil water storage state. The three control gates include the forget gate and the input gate. and output gate; according to the traditional watershed hydrological simulation theory and method, rainfall is selected as the only input factor of the LSTM flood forecast model. The input of the LSTM flood forecast model is the long-sequence rainfall information of each rainfall station in the watershed, and the input length is n+l time period; among them, n represents the input length of the previous rainfall, and the previous rainfall can be regarded as reflecting the influence of information such as soil water storage status and short-term rainfall in the basin on the output flow value of the model. Multiple n can be selected and determined according to the performance of the subsequent model. The optimal model; the output of the LSTM flood forecasting model is a sequence of discharge values equal to the confluence time of the study basin, that is, the output length is l time periods, where l is equal to the basin confluence time and l≤n; The fifth step is to generate training, validation and test sample sets; The sample length is determined according to the input and output structure (values of n and l) of the new general LSTM flood forecasting model designed in the fourth step. The length of the input rainfall sequence of each sample is equal to n+l periods, and the length of the output target flow sequence is equal to l each time period; according to the training set, validation set and test set divided in the first step, each flood will generate corresponding training, validation and test sample sets according to the method of sliding interception by time period, and each flood will generate multiple samples, Each sample consists of the input rainfall sequence P=[P _t-n+1 ,P _t-n+2 ,...,P _t ,...,P _t+l ] and the output target flow sequence Q=[Q _t+1 ,Q _t+2 ,…,Q _t ,…,Q _t+l ] constitute the input and output data pairs; among them, n is the number of periods of the input precipitation sequence of the LSTM flood forecast model, l is the number of periods of the output flow of the model, and the t th The input rainfall P _t of each time period includes the measured rainfall value of each rainfall station in the basin; The sixth step, model construction and training; The LSTM flood forecasting model uses the back-propagation over time algorithm BPTT for supervised learning training. Model construction and training are based on open source Keras and TensorFlow. The Keras framework running on the TensorFlow platform integrates a relatively mature machine learning algorithm package. The corresponding algorithm can be directly called to complete the construction and training of the LSTM flood forecasting model; 6.1) Construction of LSTM flood forecasting model: According to the number of hidden layers given in the third step, the number of hidden layer neurons and the model input and output structure designed in the fourth step, the Keras layer package is called to define the input layer and hidden layer of LSTM and the output layer to build an LSTM flood forecasting model; 6.2) LSTM flood forecasting model training: input the training and verification sample sets generated in the fifth step, set the hyperparameters, activation functions, loss functions, optimization algorithms, etc. involved in the model training process, and run the LSTM flood based on the Keras framework on the TensorFlow platform forecasting model to obtain the trained LSTM flood forecasting model; The seventh step is to determine the optimal model, and extract and analyze the flood simulation results of the basin floods; Compare and analyze the performance of the LSTM flood forecasting model under different input lengths, determine the final optimal LSTM flood forecasting model, further extract the results of flood simulation and forecast flow process in the basin, and analyze the forecasting performance of the LSTM flood forecasting model.

Images