A flood process probability prediction method based on a hybrid deep learning model

Abstract

The application discloses a flood process probability prediction method based on a hybrid deep learning model, which comprises the following steps: firstly, collecting meteorological and hydrological basic data of a research basin, establishing a conceptual model, and predicting a flood process in multiple periods; secondly, taking the flow process predicted by the conceptual model as an exogenous input, nesting a mixed density network (MDN) in an output layer of a long short-term memory (LSTM-EDE) neural network based on an exogenous input coding-decoding structure, constructing a hybrid LSTM-EDE-MDN probability prediction model, simultaneously establishing a loss function by using a maximum likelihood estimation method, training neural network parameters, and finally obtaining a conditional distribution function and a prediction interval of each prediction period, so as to quantify prediction uncertainty. The application couples the LSTM-EDE neural network with the conceptual model for predicting the flow as an exogenous input and the MDN, solves the exposure bias problem, can obtain a multi-period flood process probability prediction under the premise of considering the time correlation of an output variable, and improves the applicability, interpretability and reliability of the deep learning model.

Description

Technical Field

[0001] This invention belongs to the technical field of hydrological forecasting, specifically relating to a probabilistic flood process forecasting method based on a hybrid deep learning model. Background Technology

[0002] Hydrological forecasting is an important non-engineering measure for flood control and drought relief. Improving flood forecast accuracy and extending the lead time has always been a key technical bottleneck restricting the operation and management of reservoirs. In recent years, artificial intelligence technology has developed rapidly, leading to the emergence of deep learning models capable of effectively handling nonlinear and non-steady-state time series. Long Short-Term Memory (LSTM) neural networks are among the most representative models, achieving better forecast accuracy in multi-period flood forecasting compared to traditional artificial neural networks. Feng et al. (2019) proposed a short-term flood forecasting method for small and medium-sized rivers based on LSTM neural networks. This method can extract effective features and has high forecast accuracy, outperforming traditional support vector machine models, especially in the peak flood stage, where the forecast accuracy for peak time and peak flood value is significantly improved. However, these deep learning models lack a physical basis and have low interpretability, making them unsuitable for implementation in engineering management. Zhou et al. (2022) used forecasted flow based on a conceptual hydrological model as an additional input to a deep learning model, attempting to couple the runoff generation and confluence process during flow forecasting. Meanwhile, the forecasted flow based on the process hydrological model can supplement the input of the LSTM model in the long forecast period, alleviate the overfitting problem of the LSTM model, and improve the interpretability and forecast accuracy of the deep learning model to a certain extent.

[0003] Meanwhile, with the continuous advancement of deep learning research, encoder-decoder structures have emerged to solve sequence-to-sequence problems. The encoding process extracts important features from the input sequence and compresses them into a fixed-length intermediate vector, while the decoding process converts the intermediate vector into the target output sequence. LSTM neural networks with coupled encoder-decoder structures can pass effective features extracted in the previous time step to the next time step during both encoding and decoding. While ensuring the temporal correlation of the output variables, they can directly obtain flood forecasts with multiple lead times, exhibiting higher interpretability and applicability compared to single-output LSTM models. Wang Fan et al. (2022) proposed a flood forecasting model and method based on a deep learning framework using the encoder-decoder structure. This model can use forecasted rainfall data as model input, fully utilizing hydrological and rainfall information to improve forecast accuracy and extend the lead time.

[0004] Due to uncertainties in model inputs, parameters, and structure, such as meteorological forcing, flood forecasts inevitably involve uncertainties. The deterministic point estimation provided by deep learning models offers limited risk information for flood control decision-making. Probabilistic forecasting, however, can reflect these uncertainties and provide risk information for decision-makers. Liu et al. (2017) proposed a coupling method between ensemble precipitation forecasts and real-time probabilistic flood forecasts. By real-time correction of deterministic forecast errors and coupling ensemble precipitation forecast information with a Bayesian forecasting system based on Copula functions, they improved flood forecast accuracy and provided decision-makers with more reliable forecast intervals. Probabilistic forecasting enhances forecast value and reliability, making it essential for flood control decision-making and other related work.

[0005] In summary, there are still some shortcomings in the research of deep learning models: (1) Neural networks based on traditional encoder-decoder structures cannot learn the generation and confluence processes of conceptual hydrological models and have exposure bias problems, that is, the training process and the verification process are inconsistent, which can easily lead to low interpretability and unstable performance of deep learning models, thereby reducing forecast accuracy; (2) The output of deep learning models is mostly deterministic point estimates, which cannot provide forecast uncertainty estimates, resulting in low forecast value and credibility. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a probabilistic flood forecasting method based on a hybrid deep learning model. This method overcomes the exposure bias problem, considers the temporal correlation of output variables, and quantifies forecast uncertainty in deep learning probabilistic model methods, thereby further improving its applicability, interpretability, and reliability.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A probabilistic flood process forecasting method based on a hybrid deep learning model, characterized by the following steps:

[0009] Step 1: Collect and analyze meteorological and hydrological data, calculate the average runoff generation and confluence time of the basin, and set the forecast lead time length according to actual needs;

[0010] Step 2: Based on the data collected in Step 1, calibrate the parameters of the conceptual hydrological model, and use the calibrated conceptual hydrological model to forecast multi-period lead time flow processes;

[0011] Step 3: Using the forecast flow process of the conceptual model in Step 2 as the external input, establish a Long Short-Term Memory (LSTM-EDE) neural network, and use the output of the hidden layer of the LSTM-EDE neural network as the input of the Hybrid Density Network (MDN) to construct an LSTM-EDE-MDN model for quantifying forecast uncertainty.

[0012] Step 4: Set the activation function and hyperparameters of the LSTM-EDE-MDN model constructed in Step 3, establish the loss function to optimize the hyperparameters, and organize the input and target output variables of the LSTM-EDE-MDN model.

[0013] Step 5: Train the LSTM-EDE-MDN model based on the input and target output variables organized in Step 4, and then deduce the conditional probability distribution function of the target variable to obtain the forecast interval at a certain confidence level, thereby quantifying the uncertainty of the forecast process.

[0014] Furthermore, step 1 specifically includes:

[0015] Step 1.1: The collected meteorological and hydrological data include, but are not limited to, precipitation, temperature, evaporation, and flow at the outlet section of the basin. The time scale of the data is daily or intra-day.

[0016] Step 1.2: Divide the data into training period, validation period, and testing period data;

[0017] Step 1.3: Using the measured precipitation and flow data collected in Step 1-1, estimate the average runoff generation and confluence time of the watershed based on the correlation coefficients of precipitation and flow with different time lags. The time lag number corresponding to the largest correlation coefficient is the average runoff generation and confluence time of the watershed.

[0018] Step 1.4: Determine the forecast period length for flood forecasts based on actual flood control and other task requirements. The forecast period length shall be less than or equal to the average runoff generation and confluence time of the basin.

[0019] Furthermore, step 2 specifically includes:

[0020] Step 2-1: Select a suitable conceptual hydrological model based on the actual situation;

[0021] Step 2-2: Based on the data collected in Step 1, the SCE-UA method is used to calibrate the model parameters, verify the effectiveness of the model, and test the model performance.

[0022] Step 2-3: Use the conceptual hydrological model tested in Step 2-2 to forecast floods and obtain flow processes for multiple time periods.

[0023] Furthermore, step 3 specifically includes:

[0024] Step 3-1: Couple the LSTM neural network to the external input encoder-decoder structure (EDE) to construct the LSTM-EDE model. In the EDE structure decoding process, an interface for receiving the flow process predicted by the conceptual hydrological model is developed.

[0025] Step 3-2: Using Y as the target output variable, the hidden layer output X of the LSTM-EDE model decoding process is used as the input of the hybrid density network MDN to establish a probabilistic prediction model of LSTM-EDE-MDN hybrid deep learning. The LSTM-EDE-MDN model outputs the weights w and parameters θ of multiple kernel functions. The kernel functions are added together according to their weights w to form the conditional density function f(Y|θ,X) of the target variable Y:

[0026]

[0027]

[0028] In the formula, m is the number of kernel functions. It is the Gaussian kernel function, w i Let be the weight of the i-th kernel function.

[0029] Furthermore, step 4 specifically includes:

[0030] Step 4-1: Set the activation function and hyperparameters of the LSTM-EDE-MDN model;

[0031] Step 4-2: Construct a loss function using the maximum likelihood estimation method, and then use the loss function to optimize and adjust the hyperparameters.

[0032] Step 4-3: Measured precipitation and flow data prior to the forecast reference time are used as inputs to the LSTM-EDE-MDN model encoding process, with the input time step equal to the average runoff generation and confluence time of the watershed; the flow forecast from the conceptual hydrological model is used as input to the decoding process; the output of the LSTM-EDE-MDN model is the conditional distribution function of the target variable at each forecast time, with the output time step equal to the forecast period length; additionally, a suitable dataset for training, validating, and testing the LSTM-EDE-MDN model is compiled from the data collected in Step 1, including the inputs to the encoding and decoding processes and the target output variable.

[0033] Furthermore, in step 4-2, the constructed loss function is:

[0034]

[0035] In the formula, n is the data length of a batch;

[0036] The neural network quantifies the probability density of the target variable in the conditional distribution function output by the LSTM-EDE-MDN model through the loss function, and then adjusts the hyperparameters. The mixture density function generated by the LSTM-EDE-MDN model is f(Y│θ,X). When training the LSTM-EDE-MDN model, the adaptive moment estimation algorithm selects the parameters that maximize the probability density of the target variable Y in the log-likelihood function ln(f(Y│θ,X)).

[0037] Furthermore, step 5 specifically includes:

[0038] Step 5-1: In step 4, the corresponding input and target output variables are organized and used as the datasets for training, validating and testing the LSTM-EDE-MDN model. Multiple LSTM-EDE-MDN models are trained using the training set organized in step 4.

[0039] Step 5-2: Substitute the validation set prepared in Step 4 into the LSTM-EDE-MDN model trained in Step 5-1 to obtain the conditional probability distribution function of the target variable. With the goal of minimizing the continuous ranking probability score index, select the network parameters with the best generalization performance of the validation set from multiple LSTM-EDE-MDN models.

[0040] Step 5-3: Based on the network parameters selected in Step 5-2, the probabilistic prediction performance of the LSTM-EDE-MDN model is tested using the test set prepared in Step 4. The conditional probability distribution function of the target variable and the deterministic prediction results are derived. At the same time, the confidence level is set to obtain the prediction interval, thereby quantifying the prediction uncertainty.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] 1. The LSTM-EDE-MDN model constructed in this invention is nested with an EDE structure that can receive flow forecasts from conceptual hydrological models. It can not only learn the flow generation and confluence processes of conceptual hydrological models, but also overcome the exposure bias problem, making the training and validation processes consistent, and improving interpretability, computational efficiency and forecast accuracy.

[0043] 2. The LSTM-EDE-MDN model constructed in this invention incorporates MDN technology. It can transform deterministic point estimation into probability distribution estimation while considering the temporal correlation of output variables, thereby obtaining a forecast interval with a certain confidence level. This achieves the purpose of quantifying the uncertainty of flood forecasting, providing decision-makers with more risk information, and improving forecast value and credibility. Attached Figure Description

[0044] Figure 1This is a schematic diagram of the structure of the LSTM-EDE-MDN model in an embodiment of the present invention; wherein, (a) is the encoding process, (b) is the decoding process, and (c) is the probability prediction process;

[0045] Figure 2 This is a schematic diagram of a hybrid density network in an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of the actual flow rate, deterministic forecast, and 95% confidence forecast interval in an embodiment of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0048] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0049] The present invention will be further described below with reference to specific embodiments, but these are not intended to limit the scope of the invention.

[0050] This invention provides a method for probabilistic flood process forecasting based on a hybrid deep learning model, comprising the following steps:

[0051] Step 1 involves collecting and analyzing meteorological and hydrological data, calculating the average runoff generation and confluence time of the watershed, and setting the forecast lead time length based on actual needs. This step specifically includes:

[0052] Step 1-1: Select the study basin and collect meteorological and hydrological data, specifically including measured data on precipitation, temperature, evaporation, and flow at the basin outlet section. The time scale of the data can be daily or intra-day.

[0053] Step 1-2: Divide the data collected in Step 1-1 into training period, validation period, and testing period data;

[0054] Steps 1-3: The average runoff generation and runoff time of the basin is estimated using the precipitation-runoff relationship. The average runoff generation and runoff time of the basin is estimated based on the precipitation and runoff data collected in Step 1-1 and the correlation coefficients of precipitation and runoff with different time lags. The time lag number corresponding to the largest correlation coefficient is the average runoff generation and runoff time of the basin.

[0055] Steps 1-4: Determine the forecast period length for flood forecasts based on actual flood control and other task requirements. The forecast period length should be less than or equal to the average runoff generation and confluence time of the basin.

[0056] Step 2 involves calibrating the conceptual hydrological model parameters based on the data collected in Step 1, and then using the calibrated conceptual hydrological model to forecast multi-period lead time flow processes. This specifically includes the following steps:

[0057] Step 2-1: Select a suitable conceptual hydrological model based on the actual situation, including but not limited to the Xin'anjiang model;

[0058] Step 2-2: Based on the training period data collected in Step 1, the parameters of the conceptual hydrological model are calibrated using the SCE-UA method, the validity of the conceptual hydrological model is verified using validation period data, and the performance of the conceptual hydrological model is tested using test period data.

[0059] Steps 2-3 involve using the tested conceptual hydrological model to forecast floods and obtain flow processes for multiple time periods.

[0060] Step 3: Using the forecast flow from the conceptual model in Step 2 as external input, establish a Long Short-Term Memory (LSTM-EDE) neural network based on an external input encoder-decoder structure. Use the output of the hidden layers of the LSTM-EDE neural network as input to a Hybrid Density Network (MDN) to construct an LSTM-EDE-MDN model for quantifying forecast uncertainty. This step specifically includes:

[0061] Step 3-1: Couple the LSTM neural network to the external input encoder-decoder (EDE) structure to construct the LSTM-EDE model (see structure diagram). Figure 1 (a) and (b)). The EDE structure has developed an interface for receiving external input sequences during the decoding process, enabling it to receive predicted flow processes from conceptual hydrological models ( Figure 1 (b) Dashed line) can not only learn the runoff generation and confluence process of conceptual hydrological models, but also overcome the exposure bias problem of traditional code-decode structure, and improve interpretability, long-term forecast accuracy and computational efficiency;

[0062] Step 3-2: Using Y as the target output variable, the hidden layer output X of the LSTM-EDE model decoding process is used as the input of the hybrid density network (MDN). Figure 1 (c) Establish the proposed LSTM-EDE-MDN hybrid deep learning probabilistic prediction model. Figure 1MDN combines the LSTM-EDE model with a mixture density function, using a neural network to generate weights w and parameters for multiple kernel functions. These kernel functions are then summed according to their weights w to form a conditional density function f(Y|θ,X), where θ is the set of function parameters, to approximate the true distribution of the target variable. Flood forecast sequences are one-dimensional time series. The mixture density network uses a Gaussian kernel function, and the network output consists of the weights w, expectations μ, and variance σ of multiple kernel functions. w is normalized using a softmax function to ensure an effective discrete distribution; σ is processed using an exponential function to ensure it is a non-negative function. Given the input LSTM-EDE model hidden layer output X, the probability density function f(Y|θ,X) of the target variable Y is:

[0063]

[0064]

[0065] In the formula, m is the number of kernel functions. It is the Gaussian kernel function, w i Let be the weight of the i-th kernel function.

[0066] The commonly used kernel function is the Gaussian kernel function, and its formula is as follows:

[0067]

[0068] MDN's output variable Y f The number of elements is 3m;

[0069]

[0070] The LSTM-EDE-MDN model can transform the point estimates generated during the decoding process into probability distribution estimates, while taking into account the temporal correlation of the output variables, so as to reflect the uncertainty of the forecasting process.

[0071] Step 4: Set the activation function and hyperparameters of the LSTM-EDE-MDN model constructed in Step 3, establish the loss function using the maximum likelihood estimation method, and organize the input and target output variables of the LSTM-EDE-MDN model; this step specifically includes:

[0072] Step 4-1: Set the activation function and hyperparameters of the LSTM-EDE-MDN probabilistic prediction model. The activation function is chosen as the tanh function. Hyperparameters include the number of hidden layers in the LSTM neural network during encoding and decoding, the number of neurons in each hidden layer, and the number of kernel functions in the MDN. An n-layer LSTM neural network with m neurons is used in the encoding and decoding process, and k Gaussian kernel functions are selected for the MDN. Figure 2 A schematic diagram of an MDN containing three Gaussian kernel functions;

[0073] Step 4-2: Construct the loss function using the maximum likelihood estimation method. Unlike the loss functions of deterministic output deep learning (such as mean squared error and mean absolute error), the loss function of the LSTM-EDE-MDN probabilistic prediction model adjusts the hyperparameters by quantifying the probability density of the target variable in the network's output conditional distribution function. The mixture density function generated by the LSTM-EDE-MDN model is f(Y│θ,X). During training the LSTM-EDE-MDN model, the adaptive moment estimation (Adam) algorithm is used to select the parameters that maximize the probability density of the target variable Y in the log-likelihood function ln(f(Y│θ,X)). During backpropagation, the Adam algorithm always optimizes the neural network hyperparameters in the direction that minimizes the loss function the fastest. Therefore, the loss function is defined as:

[0074]

[0075] In the formula: n is the data length of a batch.

[0076] Step 4-3 uses measured precipitation and flow data prior to the forecast reference time as input variables for the LSTM-EDE-MDN model encoding process. The number of input time steps should equal the average runoff generation and confluence time of the watershed. The flow forecast from the conceptual hydrological model is used as the input for the decoding process, i.e., the external input sequence. The output of the LSTM-EDE-MDN model is the conditional distribution function of the target variable at each forecast time, with the number of output time steps equal to the forecast period length. Additionally, a suitable dataset for training, validating, and testing the LSTM-EDE-MDN model is compiled from the data collected in Step 1 and used as input for both the encoding and decoding processes.

[0077] Step 5: Train the LSTM-EDE-MDN model based on the input and target output variables organized in Step 4, and then deduce the conditional probability distribution function of the target variable to obtain the forecast interval at a certain confidence level, thereby quantifying the forecast uncertainty.

[0078] Step 5-1: Train multiple LSTM-EDE-MDN models using the training set prepared in Step 4. The training process includes setting Adam algorithm parameters, such as setting the learning rate to 0.001, setting the batch size and number of epochs used to train the neural network to 120 and 600 respectively, in order to quickly optimize the loss function; setting the dropout rate to 0.1 to obtain the best generalization performance, etc.

[0079] Step 5-2: Substitute the validation set data prepared in Step 4 into the multiple LSTM-EDE-MDN models trained in Step 5-1 to obtain the conditional probability distribution function of the target variable. With the goal of minimizing the Continuous Rank Probability Score (CRPS), select the network parameters with the best generalization performance on the validation set from among the multiple LSTM-EDE-MDN models. The Continuous Rank Probability Score (CRPS) evaluates the degree of fit between the conditional distribution function of the probability forecast and the true distribution of the forecast, comprehensively considering the reliability and concentration of the probability forecast. A lower CRPS value is considered to indicate better probability forecast performance; its calculation formula is:

[0080]

[0081]

[0082] In the formula: N is the number of samples, Q o,i Let represent the i-th target variable (measured flow rate). F(·) is the estimated probability distribution function, I(·) is the indicator function, and r represents the flow rate variable.

[0083] Step 5-3: Based on the network parameters selected in Step 5-2, the probabilistic prediction performance of the LSTM-EDE-MDN model is tested using the test set data organized in Step 4. This allows for the derivation of the conditional probability distribution function of the target variable and the deterministic prediction results. At the same time, a 95% confidence level is set to obtain the prediction interval, thereby quantifying the prediction uncertainty. Figure 3 This demonstrates a comparison between measured flow rates, expected flow rates calculated using the model in this embodiment, and the 95% confidence level forecast interval. From... Figure 3 It can be seen that the expected value forecast flow process calculated in this embodiment can fit the measured flow process well. At the same time, the forecast interval with a 95% confidence level can cover most of the measured flow points, indicating that the forecast interval is reasonable and reliable and can reasonably quantify the forecast uncertainty.

[0084] In summary, this invention first collects and studies basic meteorological and hydrological data of the watershed, establishes a conceptual model, and forecasts flood processes in multiple time periods. Secondly, it couples a hybrid density network (MDN) to the output layer of an LSTM-EDE neural network with the predicted flow from the conceptual model as the external input, constructing an LSTM-EDE-MDN probabilistic forecasting model. Simultaneously, it employs maximum likelihood estimation to establish a loss function and train the neural network parameters, ultimately obtaining the conditional distribution function and forecast interval for each forecast period, thus achieving probabilistic forecasting. This invention couples an LSTM-EDE neural network with the predicted flow from the conceptual model as the external input and a hybrid density function, which can solve the exposure bias problem of traditional encoder-decoder structures and learn the flow generation and confluence processes of conceptual hydrological models. Furthermore, it can obtain probabilistic forecasts of flood processes in multiple time periods while considering the temporal correlation of output variables, thereby quantifying forecast uncertainty and improving the applicability, interpretability, and reliability of deep learning models.

[0085] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the content of this specification should be included within the protection scope of the present invention.

Claims

1. A flood process probabilistic forecasting method based on a hybrid deep learning model, characterized in that, Comprising the following steps: Step 1, collecting and analyzing meteorological and hydrological data, calculating the average runoff concentration time of the basin, and setting the length of the forecast period according to the actual needs; Step 2, calibrating the conceptual hydrological model parameters according to the data collected in step 1, and using the calibrated conceptual hydrological model to forecast the flow process of multiple time periods; Step 3, taking the forecast flow process of the conceptual model in step 2 as the external input, establishing a long short-term memory LSTM-EDE neural network, taking the LSTM-EDE neural network hidden layer output as the input of the mixture density network MDN, and constructing the LSTM-EDE-MDN model for quantifying the prediction uncertainty; Comprising: Step 3-1, coupling the LSTM neural network to the external input encoding-decoding structure EDE structure, constructing the LSTM-EDE model, wherein an interface is developed in the decoding process of the EDE structure to receive the conceptual hydrological model forecast flow process; Step 3-2, taking the hidden layer output X of the LSTM-EDE model decoding process as the input of the mixture density network MDN, establishing an LSTM-EDE-MDN hybrid deep learning probability prediction model with Y as the target output variable; the LSTM-EDE-MDN model outputs the weights w and parameters of multiple kernel functions combining the kernel functions according to the weights w into the conditional density function of the target variable Y f(Y| θ,X) : ； ； wherein m is the number of kernel functions, φ( ) is a Gaussian kernel function, is the weight of the i-th kernel function; Step 4, setting the activation function and hyperparameters of the LSTM-EDE-MDN model constructed in step 3, establishing a loss function to optimize the hyperparameters, and arranging the input and target output variables of the LSTM-EDE-MDN model; comprising: Step 4-1, setting the activation function and hyperparameters of the LSTM-EDE-MDN model; Step 4-2, constructing a loss function according to the maximum likelihood estimation method, and optimizing and adjusting the hyperparameters through the loss function; Step 4-3, using the measured precipitation and flow data before the time according to which the forecast is made as the input of the encoding process of the LSTM-EDE-MDN model, and the number of input time steps is equal to the average runoff concentration time of the basin; using the conceptual hydrological model forecast flow as the input of the decoding process; the output of the LSTM-EDE-MDN model is the conditional distribution function of the target variable at each forecast time, and the number of output time steps is equal to the length of the forecast period; in addition, the data set suitable for training, verification and testing of the LSTM-EDE-MDN model is arranged from the data collected in step 1, including the input and target output variables of the encoding and decoding processes; Step 5, training the LSTM-EDE-MDN model according to the input and target output variables arranged in step 4, further deriving the conditional probability distribution function of the target variable, obtaining the forecast interval at a certain confidence level, and quantifying the uncertainty of the forecast process.

2. The flood process probabilistic forecasting method based on a hybrid deep learning model according to claim 1, wherein, Step 1 specifically comprises: Step 1.1, the collected meteorological and hydrological data includes but is not limited to precipitation, air temperature, evaporation and flow at the outlet section of the basin, and the time scale of the data is daily or intra-daily; Step 1.2, dividing the data into training, verification and testing data; Step 1.3, using the precipitation and flow data collected in step 1-1, estimating the average runoff concentration time of the basin according to the correlation coefficients of precipitation and flow at different lags, and the lag number corresponding to the maximum correlation coefficient is the average runoff concentration time of the basin; Step 1.4, according to the actual needs of the flood control task, determine the length of the flood forecast period, and the length of the forecast period is less than or equal to the average runoff concentration time of the basin. 3.The flood process probabilistic forecasting method based on the hybrid deep learning model according to claim 1, wherein, Step 2 specifically comprises: Step 2-1, select the appropriate conceptual hydrological model according to the actual situation; Step 2-2, according to the data sorted out in step 1, use SCE-UA method to calibrate model parameters, and verify the effectiveness of the model and test the performance of the model; Step 2-3, use the conceptual hydrological model tested in step 2-2 to carry out flood forecasting, and obtain the flow process of multiple time periods.

4. The flood process probabilistic forecasting method based on a hybrid deep learning model according to claim 1, wherein, In step 4-2, the loss function constructed is: ； In the formula, n is the length of a batch of data; The neural network quantifies the probability density of the target variable in the conditional distribution function output by the LSTM-EDE-MDN model through a loss function, and then adjusts the hyperparameters; the mixed density function generated by the LSTM-EDE-MDN model is f(Y│θ,X) When training the LSTM-EDE-MDN model, the parameter that makes the probability density of the target variable Y in the log-likelihood function ln (f(Y│ θ,X)) Maximum in the training set is selected through the adaptive matrix estimation algorithm.

5. The flood process probabilistic forecasting method based on a hybrid deep learning model according to claim 1, wherein, Step 5 specifically includes: Step 5-1, sort out the corresponding input and target output variables in step 4 as the data set for training, verification and testing of LSTM-EDE-MDN model, and train multiple LSTM-EDE-MDN models using the training set sorted out in step 4; Step 5-2, put the verification set sorted out in step 4 into the LSTM-EDE-MDN model trained in step 5-1, obtain the conditional probability distribution function of the target variable, and select the network parameters with the best generalization performance of the verification set from multiple LSTM-EDE-MDN models with the minimum continuous ranking probability score index as the target; Step 5-3, according to the network parameters selected in step 5-2, test the probability forecasting performance of LSTM-EDE-MDN model using the test set sorted out in step 4, derive the conditional probability distribution function and deterministic prediction result of the target variable, and set the confidence level to obtain the prediction interval, thereby realizing the quantitative prediction of uncertainty.

Images