Deep learning short-term and medium-term runoff forecasting method fusing physical mechanism

Abstract

The embodiment of the invention discloses a deep learning short-term and medium-term runoff forecasting method fusing a physical mechanism, and relates to the technical field of hydrological forecasting, and the method comprises the steps: obtaining the hydrometeorological data and weather forecasting data of a target drainage basin; inputting the hydrometeorological data and the weather forecast data into the trained multivariable mixed downscaling model to obtain model weather forecast data of a forecast period; inputting model weather forecast data and hydro meteorological data into the trained hybrid model to obtain forecast runoff data of a forecast period; determining a posterior probability density function through back calculation forecasting, and determining a probability distribution result of the actual runoff data based on the posterior probability density function; and performing flood early warning based on a probability distribution result. The method is suitable for scenes of runoff forecasting of large and medium-sized drainage basins, mountain torrent early warning of small and medium-sized drainage basins, reservoir optimization scheduling, hydropower station operation management and the like, and can provide key technical support for water safety, water resource management and water ecological protection.

Description

Technical Field

[0001] This application relates to the field of hydrological forecasting technology, and in particular to a deep learning method for short- and medium-term runoff forecasting that integrates physical mechanisms. Background Technology

[0002] Under the dual impacts of climate change and human activities, extreme hydrological events are occurring more frequently, and the nonlinearity and uncertainty of hydrological processes have significantly increased, posing higher demands on regional flood control security, real-time water resource allocation, and emergency management. Improving the accuracy and reliability of short- and medium-term runoff forecasts has significant scientific and practical value.

[0003] Currently, short-term runoff forecasting methods mainly fall into three categories: data-driven methods, process-driven methods, and hybrid modeling methods. Data-driven methods rely on historical meteorological and hydrological data, using statistical learning or machine learning techniques to uncover nonlinear relationships in runoff formation, thereby forecasting future short-term runoff. In recent years, machine learning, especially deep learning (DL), has been widely used in data-driven runoff forecasting due to its powerful feature extraction and sequence modeling capabilities. However, these methods lack clear physical mechanisms and are highly dependent on training samples. In areas with limited data or when facing extreme hydrological events, they often suffer from decreased generalization ability and unstable forecast results due to a lack of clear mechanisms and insufficient samples. Process-driven methods are based on distributed or lumped hydrological models, combined with numerical weather prediction products for runoff forecasting. However, their forecasting effectiveness is significantly affected by the accuracy of the input meteorological forecast, especially in rapidly changing weather systems such as severe convective weather, where error accumulation and propagation problems are particularly prominent. To overcome the limitations of single methods, related technologies can adopt a hybrid modeling approach, combining data-driven and process-driven models to fully leverage their respective advantages and improve forecast accuracy and robustness. A typical practice is to simultaneously input hydrological model output and real-time monitoring data into a machine learning model, which can improve forecast performance to some extent. However, current hybrid modeling methods that integrate physical mechanisms and deep learning mostly focus on introducing multi-source data or post-processing the output results, rather than achieving an organic coupling between physical constraints and data-driven processes within the model structure. This results in limited forecasting capabilities for extreme events or unobserved watersheds, and the generalization and robustness remain unsatisfactory.

[0004] Therefore, there is currently a lack of a short- to medium-term runoff forecasting method that deeply integrates physical mechanisms and deep learning to improve the accuracy and reliability of short- to medium-term runoff forecasting. Summary of the Invention

[0005] This application provides a deep learning-based short- to medium-term runoff forecasting method that integrates physical mechanisms to address the shortcomings of the aforementioned related technologies. The technical solution is as follows: In a first aspect, embodiments of this application provide a deep learning-based short- to medium-term runoff forecasting method that integrates physical mechanisms, including: Acquire hydrological and meteorological data and weather forecast data for the target watershed; The hydrological and meteorological data and the meteorological forecast data are input into a trained multivariate mixed downscaling model to obtain the model meteorological forecast data for the lead time period. The model meteorological forecast data and the hydrometeorological data are input into the trained hybrid model and processed to obtain the forecast runoff data for the foreseeable period. The posterior probability density function corresponding to the forecast runoff data for the forecast period is determined by back-calculation forecast, and the probability distribution result of the actual runoff data within the forecast period is determined based on the posterior probability density function. Flood warnings are issued based on the probability distribution results.

[0006] In one alternative embodiment of the first aspect, after acquiring the hydrological and meteorological data and meteorological forecast data of the target watershed, the method further includes: Outlier values ​​are detected in various types of data based on a preset outlier detection method, and then deleted or corrected. Identify the missing values ​​in the various types of data obtained, and fill in the corresponding missing values ​​through interpolation; Perform time matching on various types of data and align the data to a unified timestamp based on the corresponding time. Normalize each type of data after it has been aligned to a unified timestamp, and output the normalized data.

[0007] In one alternative embodiment of the first aspect, the step of inputting the hydrological and meteorological data and the meteorological forecast data into a trained multivariate mixture downscaling model and processing them to obtain the model's meteorological forecast data for the lead time includes: The weather forecast data is processed by the convolutional neural network of the multivariate hybrid downscaling model to obtain downscaled weather forecast data; The downscaled meteorological forecast data is corrected for bias by combining the hydrological and meteorological data to obtain bias-corrected meteorological forecast data; The model weather forecast data is obtained based on the bias-corrected weather forecast data.

[0008] In one alternative to the first aspect, the hybrid model includes the Xin'anjiang model and a long short-term memory network; The process of inputting the model meteorological forecast data and the hydrometeorological data into the trained hybrid model and processing them to obtain the forecast runoff data for the foreseeable period includes: The hydrological and meteorological data are processed using the Xin'anjiang model to obtain the hydrological state variables of the target watershed; The hydrological and meteorological data, the model meteorological forecast data, and the hydrological state variables are processed by the long short-term memory network to obtain the forecast runoff data for the forecast period.

[0009] In one alternative embodiment of the first aspect, the method further includes: Acquire satellite remote sensing data of the target watershed; After obtaining the forecast runoff data for the foreseeable period through the processing, the method further includes: The satellite remote sensing data, the forecast runoff data for the forecast period, and the hydro-meteorological data are input into the trained runoff correction model so that the runoff correction model can correct the forecast runoff data for the forecast period in real time based on the hydro-meteorological data and the satellite remote sensing data, thereby obtaining the corrected forecast runoff data.

[0010] In one alternative embodiment of the first aspect, determining the posterior probability density function corresponding to the forecast runoff data for the forecast period through back-calculation forecasting includes: Select hydrological and meteorological data of the target watershed in a preset historical period, select the forecast time within the preset historical period, and determine the back-calculation forecast period after the forecast time within the preset historical period. Determine the back-calculated forecast runoff data for the back-calculated forecast period, and determine the actual runoff data for the back-calculated forecast period and the actual runoff data at the forecast time based on the hydrological and meteorological data for the preset historical period; The distribution parameters of the back-calculated forecast runoff data, the actual runoff data, and the actual runoff data at the forecast time are determined respectively, and the corresponding probability density functions are obtained by fitting the distribution parameters. A two-dimensional joint distribution of connection functions is constructed based on the actual runoff data at the forecast time and the actual runoff data. A three-dimensional joint distribution of connection functions is constructed based on the back-calculated forecast runoff data, the actual runoff data at the forecast time and the actual runoff data. The posterior probability density function of the actual runoff data is constructed based on the probability density function, the joint distribution of the two-dimensional connection function, and the joint distribution of the three-dimensional connection function. The posterior probability density function of the actual runoff data is used as the posterior probability density function of the predicted runoff data for the forecast period.

[0011] In one alternative embodiment of the first aspect, determining the probability distribution of the actual runoff data within the forecast period based on the posterior probability density function includes: Based on the hydrological and meteorological data, the actual runoff data at the moment before the forecast period is determined. The predicted runoff data for the forecast period and the posterior probability density function are combined to perform a probability forecast of the actual runoff data for the forecast period, thereby obtaining the probability distribution result of the actual runoff data for the forecast period. The flood warning based on the probability distribution results includes: Based on the probability distribution results, the probability that the actual runoff value at each time point within the forecast period is greater than the runoff threshold of different levels at the corresponding time point is determined. If the probability is greater than the probability threshold for the corresponding level, a flood warning for the current level will be issued.

[0012] Secondly, embodiments of this application also provide a deep learning short- to medium-term runoff forecasting device that integrates physical mechanisms, comprising: The data acquisition unit is used to acquire hydrological and meteorological data and weather forecast data for the target watershed. The forecast data processing unit is used to input the hydro-meteorological data and the meteorological forecast data into a trained multivariate mixed downscaling model to obtain the model meteorological forecast data for the lead time period. The runoff forecasting unit is used to input the model meteorological forecast data and the hydrometeorological data into the trained hybrid model and process them to obtain the forecast runoff data for the foreseeable period. The back-calculation forecast unit is used to determine the posterior probability density function corresponding to the forecast runoff data in the forecast period through back-calculation forecast, and to determine the probability distribution result of the actual runoff data in the forecast period based on the posterior probability density function; An early warning unit is used to issue flood warnings based on the probability distribution results.

[0013] Thirdly, embodiments of this application also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method provided by the first aspect or any implementation thereof of the embodiments of this application.

[0014] Fourthly, this application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method provided by the first aspect of the embodiments of this application or any implementation thereof.

[0015] The beneficial effects of the technical solutions provided in some embodiments of this application include at least the following: (1) By integrating hydrological and meteorological data and meteorological forecast data, model meteorological forecast data with reasonable spatial distribution and strong physical consistency can be generated, thereby improving the quality and reliability of meteorological forecast data from the source.

[0016] (2) By constructing a hybrid model that integrates hydrological physical model and machine learning, the advantages of mechanism model and data-driven model can be organically combined. It retains the ability of hydrological model to characterize physical processes and makes full use of the powerful nonlinear fitting and pattern recognition capabilities of data-driven model, thereby improving the forecast accuracy and stability of runoff of different magnitudes (especially extreme floods) and obtaining more accurate forecast runoff data for the lead time.

[0017] (3) By back-calculating the forecast, the posterior probability density function corresponding to the forecast runoff data is constructed, thereby quantitatively characterizing the forecast uncertainty and providing the maximum possible runoff forecast value and the corresponding probability, which helps decision-makers to conduct comprehensive risk assessment and formulate graded emergency plans, and promotes the effective connection between forecasting and decision-making.

[0018] (4) It is applicable to scenarios such as runoff forecasting in large and medium-sized watersheds, flash flood warning in small and medium-sized watersheds, reservoir optimization scheduling and hydropower station operation and management, and can provide key technical support for water security, water resources management and water ecological protection. It not only has the opportunity to improve the accuracy and reliability of short and medium-term runoff forecasts, but will also provide a new scientific and technological basis for dealing with data-free areas and highly uncertain hydrological events. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a deep learning-based short-to-medium-term runoff forecasting method that integrates physical mechanisms, as provided in an embodiment of this application. Figure 2 This is a schematic diagram of the structure of a CNN-based deep learning downscaling model provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of the multivariate mixture downscaling model DL-BC provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of the hybrid model XAJ-LSTM provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of a deep learning short-to-medium-term runoff forecasting device that integrates physical mechanisms, provided in an embodiment of this application. Figure 6 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or modules is not limited to the steps or modules listed, but may optionally include steps or modules not listed, or may optionally include other steps or modules inherent to such process, method, product, or apparatus.

[0023] It should be noted that the terms "first" and "second" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects. It is understood that "first" and "second" can be interchanged in a specific order or sequence where permitted. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in an order other than those described or illustrated herein.

[0024] The present application will now be described in detail with reference to specific embodiments.

[0025] Next, combine Figure 1 This paper introduces a deep learning-based short- to medium-term runoff forecasting method that integrates physical mechanisms, provided by embodiments of this application. For details, please refer to... Figure 1 , Figure 1 This illustration shows a flowchart of a deep learning-based short-to-medium-term runoff forecasting method incorporating physical mechanisms, provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps: S101, acquire hydrological and meteorological data and weather forecast data for the target watershed; S102, input the hydrological and meteorological data and the meteorological forecast data into the trained multivariate mixed downscaling model, and process them to obtain the model meteorological forecast data for the lead time period; S103, Input the model meteorological forecast data and the hydrometeorological data into the trained hybrid model, and process them to obtain the forecast runoff data for the foreseeable period; S104, determine the posterior probability density function corresponding to the forecast runoff data in the forecast period by back-calculation forecast, and determine the probability distribution result of the actual runoff data in the forecast period based on the posterior probability density function; S105, Flood warning is issued based on the probability distribution results.

[0026] Specifically, in S101, the target watershed can be any designated research area. Hydrological and meteorological data and weather forecast data for the target area can be obtained through official institutions and open-source data. The hydrological and meteorological data includes the observed meteorological data and observed flow data of the target watershed before the forecast period. The observed meteorological data includes, but is not limited to, daily precipitation, evapotranspiration, daily maximum and minimum temperatures, etc. The observed flow data includes the daily runoff observation data of the watershed outlet section of the target watershed, that is, the observed actual runoff data. For example, the period from 2000 to 2018 can be selected as the historical period, and the period after 2018 can be selected as the forecast period. The daily inflow data of Gaobazhou Reservoir in the Qingjiang River Basin during the historical period can be selected. This data can be used as the observed flow data of the outlet section of the basin. The observed meteorological data of each meteorological station in the upstream basin of the reservoir during the corresponding period can be obtained, including daily precipitation, evapotranspiration, and daily maximum and minimum temperature data. Specifically, the meteorological forecast data includes historical meteorological forecast data prior to the forecast period and future meteorological forecast data for the forecast period, specifically including daily precipitation, maximum temperature and minimum temperature forecast data covering the output of numerical weather prediction models for the corresponding period.

[0027] Specifically, the forecast period can be set to 1-7 days as needed, and this application embodiment does not limit this.

[0028] For example, historical forecast datasets and forecast data for the foreseeable period can be selected from the European Centre for Medium-Range Weather Forecasts (ECMWF) weather forecast model.

[0029] In some embodiments, after S101, consistency checks and quality control can also be performed on the multi-data information obtained in S101, specifically including: Outlier detection and handling: Based on a preset outlier detection method, outliers in various types of data are detected, and then deleted or corrected. For example, outliers can be identified through statistical methods (such as box plots) or physical thresholds, and then corrected or removed. Identify and handle missing values: Determine the missing values ​​in the acquired data and fill them in using interpolation; for example, depending on the data type and missing mechanism, imputation can be performed using the mean / median, time series interpolation (such as linear interpolation, spline interpolation), or spatial interpolation methods based on neighboring sites. Unified timestamp and data alignment: Time matching is performed on various types of data, and the data is aligned to a unified timestamp based on the corresponding time. Normalization processing: Normalize the various types of data aligned to a unified timestamp, and output the normalized data. Specifically, in order to eliminate the influence of units and improve the training effect and convergence speed of subsequent machine learning models, normalize various meteorological and hydrological factors. For example, use Min-Max standardization or Z-score standardization for continuous variables (such as precipitation, temperature, runoff, etc.). Normalization parameters can be constructed for forecast data and observation data respectively, and the corresponding parameters can be applied for real-time processing during the forecast stage.

[0030] In some embodiments, in S102, the observed meteorological data from the hydro-meteorological data obtained in S101 can be selected as the observed data, and the historical meteorological forecast data and the future meteorological forecast data in the forecast period can be selected as the meteorological forecast data and input into the multivariate mixed downscaling model.

[0031] Specifically, firstly, historical weather forecast data is processed using a Convolutional Neural Network (CNN). and future weather forecast data The data were processed separately to obtain the corresponding downscaling results, i.e., the downscaled historical weather forecast data. and downscaled future weather forecast data .

[0032] The convolutional neural network consists of four standard convolutional blocks, each containing two convolutional layers (3×3 convolutional kernels) and a batch normalization layer. Upsampling is achieved through sub-pixel convolution (SPC), with each upsampling block consisting of one convolutional layer and one SPC layer, and using parameterized ReLU as the activation function.

[0033] Convolutional neural networks can be used to downscale and upsample low-resolution data into high-resolution data.

[0034] Furthermore, by combining a multivariate mixture downscaling model with observed meteorological data, the downscaled historical weather forecast data were analyzed. and downscaled future weather forecast data Bias correction is performed to further reduce the bias of variables such as precipitation and temperature during the forecast period and to reconstruct the dependence between various meteorological variables. The model meteorological forecast data is obtained based on the bias-corrected meteorological forecast data.

[0035] The deviation correction process specifically includes: S1021, through a random orthogonal rotation matrix Process the observed meteorological data separately Downscaled historical weather forecast data and the downscaled future weather forecast data Apply the formula: ; ; ; in, Let N be a uniformly distributed random orthogonal rotation matrix, where N represents the observed meteorological data. The number of meteorological variables in China This represents the number of iterations.

[0036] S1022, the observed meteorological data processed by the aforementioned random orthogonal rotation matrix Historical weather forecast data, processed by the random orthogonal rotation matrix using the quantile mapping method, is used as the target data. and future weather forecast data Perform bias correction to obtain bias-corrected historical weather forecast data. and future weather forecast data .

[0037] In some embodiments, for precipitation variables, the quantile mapping method includes two steps: first, performing rainy day frequency correction on the simulated data based on precipitation data; and second, performing quantile correction on this basis. For non-precipitation variables, quantile correction is performed directly. The formula for quantile correction includes: ; ; in, This represents the i-th variable in the weather forecast data. , and These are the cumulative distribution functions of simulated meteorological data, observed meteorological data, and meteorological forecast data, respectively.

[0038] S1023, through the inverse matrix of the random orthogonal rotation matrix Processing the historical weather forecast data after deviation correction and future weather forecast data : ; ; ; In some embodiments, to solve the problem that physical consistency is broken during independent calibration in S1022, for example, the situation where Tmax < Tmin occurs, based on quantile calibration, in the embodiments of the present application, after each rotation ends, a hard constraint of Tmax > Tmin is imposed on and . Taking physical consistency as an inherent optimization condition to ensure the rationality of the iterative path and the final solution, compared with only physically correcting the final calibration result, this method helps to guide the calibration process towards a physically reasonable and statistically optimal result, avoiding the introduction of new statistical biases.

[0039] S1024, if the multivariate distribution of the historical meteorological forecast data after inverse matrix processing is consistent with the multivariate distribution of the observed meteorological data after processing by the stochastic orthogonal rotation matrix , then the future meteorological forecast data after inverse matrix processing is output as the model meteorological forecast data for the prediction period.

[0040] Otherwise, go to the step of S1021.

[0041] In some embodiments, the structural schematic diagram of the deep learning downscaling model based on CNN is as shown in Figure 2 . The CN05.1 observed meteorological data and ECMWF meteorological forecast data in the historical period (2000 - 2018) can be selected to train and validate the model. Specifically, 2000 - 2012 can be selected as the training period of the model, 2013 - 2018 as the validation period of the model, and the Adam optimizer is used to train the model with the mean absolute error (MAE) as the loss function, so that the model can achieve the function of downscaling processing; after the CNN model training is completed, the ECMWF forecast data in the historical period and the ECMWF meteorological forecasts in the future period (2019 - 2024) are respectively input into the model to obtain the historical and future meteorological forecast downscaling results 、 . The structural schematic diagram of the multivariate hybrid downscaling model DL - BC that combines bias correction and deep learning is as shown in Figure 3 . The CN05.1 observed meteorological data in the historical period (2000 - 2018), the historical period ECMWF downscaling results and the future period ECMWF downscaling results can be selected as the input data of the DL - BC model to further correct the bias of the downscaling results of ECMWF in the historical and future periods.

[0042] In some embodiments, the hybrid model XAJ-LSTM in S103 includes the Xin'anjiang Model (XAJ) and a Long Short-Term Memory (LSTM) network. In step S103, the model meteorological forecast data and the hydrometeorological data are input into the trained hybrid model, and processed to obtain the forecast runoff data for the lead time period, specifically including: S1031, The hydrological and meteorological data are processed by the Xin'anjiang model to obtain the hydrological state variables of the target watershed.

[0043] Specifically, precipitation, potential evapotranspiration, and temperature data from hydrological and meteorological data, as well as daily runoff observation data from the basin outlet section, can be used as model inputs. The parameters of the Xin'anjiang model can be optimized using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) algorithm. This algorithm minimizes the error between simulated runoff and observed runoff by continuously adjusting the model parameters, thereby achieving model parameter calibration. This application does not limit the specific implementation of this method.

[0044] Related technologies generally use the Xin'anjiang model to process hydrological and meteorological data to obtain predicted runoff. In this application, the intermediate data obtained by processing hydrological and meteorological data using the Xin'anjiang model are used to obtain the hydrological state variables of the target watershed, including but not limited to surface runoff, groundwater runoff, interflow, and water storage in the three soil layers.

[0045] S1032, the hydrological and meteorological data, the model meteorological forecast data, and the hydrological state variables are processed through the long short-term memory network to obtain the forecast runoff data for the forecast period.

[0046] Specifically, during the training of the hybrid model, historical data on inflow into the Gaobazhou Reservoir (2000-2018) and meteorological variables observed at hydrological stations (daily precipitation, maximum and minimum temperatures) can be used for the calibration and validation of the XAJ model.

[0047] Specifically, the period from 2000 to 2012 was designated as the calibration period for the XAJ model, and the period from 2013 to 2018 was designated as the validation period. The NSE index was used to evaluate the model's ability to simulate runoff, and the requirement was met when the NSE index during the model validation period was greater than 0.8.

[0048] Specifically, after the XAJ model meets the standards, the historical watershed hydrological state variables (including surface runoff, groundwater runoff, interflow, and three-layer soil water storage) are output as input data for the hybrid model XAJ-LSTM. A schematic diagram of the structure of the hybrid model XAJ-LSTM, which combines a physics-based hydrological model and a machine learning model, is shown below. Figure 4 As shown, inflow runoff data from 2000 to 2012 can be selected as target data, and observed meteorological data from 2000 to 2012 and watershed hydrological state variables output by XAJ can be used as input data to train the LSTM model. The period from 2013 to 2018 is selected as the validation period, and NSE is used to evaluate the runoff simulation effect of the hybrid model XAJ-LSTM.

[0049] Understandably, the trained LSTM model can obtain forecast runoff data for the foreseeable period based on hydrological and meteorological data, model meteorological forecast data, and hydrological state variables.

[0050] Optionally, after obtaining the forecast runoff data for the foreseeable period in S103, the forecast runoff data for the foreseeable period can be corrected in real time, specifically including: S201, acquire satellite remote sensing data of the target watershed.

[0051] Specifically, satellite remote sensing data includes meteorological and hydrological information of the target watershed, such as remote sensing image data of the target watershed in historical periods, and high spatiotemporal resolution satellite inversion precipitation products, etc. This application does not limit this.

[0052] After obtaining the forecast runoff data for the foreseeable period through the processing, the method further includes: S202, the satellite remote sensing data, the forecast runoff data for the forecast period, and the runoff observation data from the hydrological and meteorological data are input into the trained runoff correction model, so that the runoff correction model can perform real-time correction on the forecast runoff data for the forecast period based on the hydrological and meteorological data and the satellite remote sensing data, and obtain the corrected forecast runoff data.

[0053] Specifically, the runoff correction model can use runoff observation data as the target variable, and satellite remote sensing data and forecast runoff data for each forecast period as input data for training, respectively, to construct a real-time LSTM-based forecast runoff correction model with a forecast period of 1-7 days. k (k=1-7). After training, the runoff correction model can assimilate measured hydrological data and meteorological and hydrological information monitored by satellite remote sensing in real time.

[0054] In this way, a real-time correction model based on deep learning dynamically assimilates real-time monitoring and remote sensing data to achieve real-time deviation correction of hydrological model forecast results. This can promptly correct forecast drift caused by model structure or input errors, thereby enhancing the adaptability and robustness of the forecast system.

[0055] In some embodiments, in S104, the posterior probability density function corresponding to the forecast runoff data for the forecast period can be determined by back-calculation forecasting, specifically including: S1041, Select hydrological and meteorological data of the target watershed in a preset historical period, select the forecast time within the preset historical period, and determine the back-calculation forecast period after the forecast time within the preset historical period.

[0056] S1042, determine the back-calculated forecast runoff data for the back-calculated forecast period, and determine the actual runoff data for the back-calculated forecast period and the actual runoff data at the forecast time based on the hydrological and meteorological data of the preset historical period.

[0057] For example, 2012-2018 can be selected as the preset historical period for back-calculation forecasting. Hydrological and meteorological data and meteorological forecast data within the preset historical period can be easily obtained. Multiple forecast times can be selected within 2012-2018. Based on the given forecast period length, the time interval after the forecast time within the preset historical period can be selected as the back-calculation forecast period.

[0058] Specifically, the back-calculated forecast runoff data for each back-calculated forecast period can be obtained by processing hydrological and meteorological data and meteorological forecast data within a preset historical period through the S102-S103 model. The actual runoff data for the corresponding period can be directly extracted based on the hydrological and meteorological data within the preset historical period.

[0059] S1043, determine the distribution parameters of the back-calculated forecast runoff data, the actual runoff data, and the actual runoff data at the forecast time, and obtain their respective probability density functions based on the distribution parameters.

[0060] Specifically, taking a 1-7 day forecast period as an example, the fitting method can be used to estimate the back-calculated forecast runoff data for the 1-7 day forecast period. Actual runoff data and actual runoff data at the forecast time The P-III distribution parameters were determined, and the Kolmogorov-Smirnov (KS) test was used to test the fit of these distributions, and their probability density functions were analyzed. for: ; in, , and These are shape, scale, and position parameters, respectively. This represents the gamma function, which serves a normalization purpose.

[0061] S1044, based on the actual runoff data at the predicted time. and the actual runoff data The joint distribution of the two-dimensional connection function (i.e., the two-dimensional time-varying Copula joint distribution) is constructed.

[0062] Simultaneously, runoff data can be back-calculated and predicted. Actual runoff data at the predicted time and the actual runoff data The joint distribution of three-dimensional connection functions (i.e., the static three-dimensional Copula joint distribution) is constructed.

[0063] Specifically, based on the actual runoff data obtained from daily observations within the selected preset historical period, the rolling time window method is used, with the annual accumulated days (doy) as the time-varying driving factor t, and the correlation time-varying parameters of the Copula function are estimated using maximum likelihood estimation. t .

[0064] Specifically, the parameter values ​​of the Copula function can be calculated using the maximum likelihood method by using the actual runoff data obtained from daily observations within the selected preset historical period, the back-calculated forecast runoff data for each forecast period, and the actual runoff data.

[0065] Understandably, Copula is used to describe the dependency structure among multiple random variables, and it can connect marginal distributions to form a joint distribution.

[0066] Specifically, the fit of two-dimensional and three-dimensional Copula functions can be evaluated using RMSE. When calculating the RMSE of a two-dimensional Copula function, the parameters need to be considered. t It varies over time; when calculating the RMSE of a three-dimensional Copula function, it is necessary to distinguish between symmetric and asymmetric types, and select the type with the smaller RMSE value as the connection function, thereby constructing... , and The joint distribution of .

[0067] S1045, the posterior probability density function of the actual runoff data is constructed based on the probability density function, the joint distribution of the two-dimensional connection function, and the joint distribution of the three-dimensional connection function.

[0068] Specifically, this can be achieved by back-calculating the actual runoff data for the forecast period k. The marginal distribution function and its corresponding probability density function Joint distribution with two-dimensional connection function Determine the actual runoff data Time-varying prior density function : ; in, Actual runoff data The actual value in, Actual runoff data at the forecast time The actual value, For uncertain variables, they refer to respectively here. and .

[0069] Specifically, the distribution can be jointly distributed using a two-dimensional connection function. Joint distribution of three-dimensional connection functions And the back-calculated forecast runoff data at the forecast period k. The marginal distribution function and its corresponding probability density function Determine the actual runoff data Time-varying likelihood function : ; in, To back-calculate forecast runoff data The actual value in, "Uncertain variable" here refers to .

[0070] The actual runoff data were calculated based on Bayes' theorem. The posterior probability density function: ; in, for The time-varying posterior probability density function, for The time-varying prior probability density function, for The time-varying likelihood function.

[0071] S1046, the posterior probability density function of the actual runoff data is used as the posterior probability density function of the predicted runoff data for the forecast period.

[0072] In some embodiments, S104 further determines the probability distribution of the actual runoff data within the forecast period based on the posterior probability density function, including: Based on the hydrological and meteorological data, the actual runoff data at the moment before the forecast period is determined. The predicted runoff data for the forecast period and the posterior probability density function are combined to perform a probability forecast of the actual runoff data for the forecast period, thereby obtaining the probability distribution result of the actual runoff data for the forecast period. The flood warning based on the probability distribution results includes: Based on the probability distribution results, the probability that the actual runoff value at each time point within the forecast period is greater than the runoff threshold of different levels at the corresponding time point is determined. Specifically, the mean of the posterior probability (50th percentile) can be used as the prediction result, with a given significance level. =0.1, calculate the lower and upper confidence limits of the 90% flow forecast interval for the actual runoff data in the forecast period. For example, different levels of runoff thresholds can be designed based on the flood control standards of Gaobazhou Reservoir. The actual runoff values ​​at each time point within the forecast period can be determined based on the probability distribution results. The probability P of exceeding different levels of runoff thresholds > ), subscript This is the label for the corresponding level.

[0073] Specifically, flood warnings of corresponding levels can be issued based on the calculated probability of exceeding the standard and preset warning rules. For example, in the case of probability P( > (Greater than the corresponding level) If the probability threshold is met, it indicates that the current level will be issued. Flood warning.

[0074] The following are apparatus embodiments of this application, which can be used to execute the method embodiments of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the method embodiments of this application.

[0075] Please see below. Figure 5 The image shows a schematic diagram of a deep learning short-to-medium-term runoff forecasting device that integrates physical mechanisms, provided as an exemplary embodiment of this application. The device includes: The data acquisition unit is used to acquire hydrological and meteorological data and weather forecast data for the target watershed. The forecast data processing unit is used to input the hydro-meteorological data and the meteorological forecast data into a trained multivariate mixed downscaling model to obtain the model meteorological forecast data for the lead time period. The runoff forecasting unit is used to input the model meteorological forecast data and the hydrometeorological data into the trained hybrid model and process them to obtain the forecast runoff data for the foreseeable period. The back-calculation forecast unit is used to determine the posterior probability density function corresponding to the forecast runoff data in the forecast period through back-calculation forecast, and to determine the probability distribution result of the actual runoff data in the forecast period based on the posterior probability density function; An early warning unit is used to issue flood warnings based on the probability distribution results.

[0076] It should be noted that the apparatus provided in the above embodiments is only illustrated by the division of the functional modules described above when executing the deep learning short- and medium-term runoff forecasting method that integrates physical mechanisms. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus provided in the above embodiments and the embodiments of the deep learning short- and medium-term runoff forecasting method that integrates physical mechanisms belong to the same concept, and the implementation process is detailed in the method embodiments, which will not be repeated here.

[0077] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described above.

[0078] Please see Figure 6 This is a structural block diagram of an electronic device provided in an embodiment of this application.

[0079] like Figure 6 As shown, the electronic device 600 includes a processor 601 and a memory 602.

[0080] In this embodiment, the processor 601 is the control center of the computer system, and can be a processor of a physical machine or a processor of a virtual machine. The processor 601 may include one or more processing cores, such as a 4-core processor or an 8-core processor. The processor 601 can be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array).

[0081] Processor 601 may also include a main processor and a coprocessor. The main processor is a processor used to process data in the wake-up state, also known as a CPU (Central Processing Unit). The coprocessor is a low-power processor used to process data in the standby state.

[0082] Memory 602 may include one or more computer-readable storage media, which may be non-transitory. Memory 602 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments of this application, the non-transitory computer-readable storage media in memory 602 is used to store at least one instruction, which is executed by processor 601 to implement the method in the embodiments of this application.

[0083] In some embodiments, the electronic device 600 further includes a peripheral device interface 603 and at least one peripheral device 604. The processor 601, memory 602, and peripheral device interface 603 can be connected via a bus or signal line. Each peripheral device 604 can be connected to the peripheral device interface 603 via a bus, signal line, or circuit board. Specifically, the peripheral device 604 includes: a display screen, a camera, and audio circuitry. The peripheral device interface 603 can be used to connect at least one I / O (Input / Output) related peripheral device to the processor 601 and memory 602.

[0084] In some embodiments of this application, the processor 601, memory 602, and peripheral device interface 603 are integrated on the same chip or circuit board; in other embodiments of this application, any one or two of the processor 601, memory 602, and peripheral device interface 603 can be implemented on separate chips or circuit boards. This application does not specifically limit the implementation in this regard.

[0085] The block diagram of the electronic device shown in the embodiments of this application does not constitute a limitation on the electronic device 600. The electronic device 600 may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0086] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the methods in any of the foregoing embodiments. The computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, microdrives, as well as magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0087] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of software products. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method of short-medium term runoff forecasting using deep learning with fusion of physical mechanisms, characterized in that, The method comprises the following steps: obtaining hydro-meteorological data and meteorological forecast data of a target basin; inputting the hydro-meteorological data and the meteorological forecast data into a trained multivariate hybrid downscaling model to obtain model meteorological forecast data of a prediction period; inputting the model meteorological forecast data and the hydro-meteorological data into a trained hybrid model to obtain forecast runoff data of the prediction period; determining a posterior probability density function corresponding to the forecast runoff data of the prediction period through back-calculation prediction, and determining a probability distribution result of actual runoff data in the prediction period based on the posterior probability density function; performing flood warning based on the probability distribution result.

2. The method according to claim 1, wherein, After obtaining the hydro-meteorological data and the meteorological forecast data of the target basin, the method further comprises the following steps: detecting abnormal values in the obtained various types of data based on a preset abnormal value detection method, and deleting or correcting the abnormal values; determining missing values in the obtained various types of data, and filling the corresponding missing values through interpolation; time-matching the various types of data, and aligning the data to a unified timestamp based on the time of the data; respectively normalizing the various types of data after being aligned to the unified timestamp, and outputting the normalized data.

3. The method according to claim 1 or 2, wherein, The step of inputting the hydro-meteorological data and the meteorological forecast data into the trained multivariate hybrid downscaling model to obtain model meteorological forecast data of a prediction period comprises the following steps: processing the meteorological forecast data through a convolutional neural network of the multivariate hybrid downscaling model to obtain down-scaled meteorological forecast data; performing bias correction on the down-scaled meteorological forecast data in combination with the hydro-meteorological data to obtain bias-corrected meteorological forecast data; obtaining the model meteorological forecast data based on the bias-corrected meteorological forecast data.

4. The method according to claim 3, wherein, The hybrid model comprises a Xin'anjiang model and a long short-term memory network. The step of inputting the model meteorological forecast data and the hydro-meteorological data into the trained hybrid model to obtain forecast runoff data of a prediction period comprises the following steps: processing the hydro-meteorological data through the Xin'anjiang model to obtain hydrological state variables of the target basin; processing the hydro-meteorological data, the model meteorological forecast data, and the hydrological state variables through the long short-term memory network to obtain forecast runoff data of the prediction period.

5. The method of claim 4, wherein the method is characterized by, The method further comprises the following steps: obtaining satellite remote sensing data of the target basin; After obtaining the forecast runoff data of the prediction period, the method further comprises the following steps: inputting the satellite remote sensing data, the forecast runoff data of the prediction period, and the hydro-meteorological data into a trained runoff correction model, so that the runoff correction model corrects the forecast runoff data of the prediction period in real time based on the hydro-meteorological data and the satellite remote sensing data to obtain corrected forecast runoff data.

6. The method of claim 4, wherein the method is characterized by, The step of determining a posterior probability density function corresponding to the forecast runoff data of the prediction period through back-calculation prediction comprises the following steps: selecting hydro-meteorological data of the target basin in a preset historical period, selecting a forecast time in the preset historical period, and determining a back-calculation prediction period after the forecast time in the preset historical period; determining the back-calculation forecast runoff data of the prediction period, determining the actual runoff data of the prediction period and the actual runoff data at the forecast time according to the hydro-meteorological data of the preset historical period; respectively determining the distribution parameters of the back-calculation forecast runoff data, the actual runoff data and the actual runoff data at the forecast time, and fitting the respective corresponding probability density functions based on the distribution parameters; constructing a two-dimensional connection function joint distribution based on the actual runoff data at the forecast time and the actual runoff data, and constructing a three-dimensional connection function joint distribution based on the back-calculation forecast runoff data, the actual runoff data at the forecast time and the actual runoff data; constructing the posterior probability density function of the actual runoff data based on the probability density functions, the two-dimensional connection function joint distribution and the three-dimensional connection function joint distribution; taking the posterior probability density function of the actual runoff data as the posterior probability density function corresponding to the forecast runoff data of the prediction period.

7. The method according to claim 6, wherein, The probability distribution result of the actual runoff data in the prediction period based on the posterior probability density function includes: determining the actual runoff data at the previous time of the prediction period based on the hydro-meteorological data, combining the forecast runoff data of the prediction period and the posterior probability density function to probabilistically forecast the actual runoff data of the prediction period, and obtaining the probability distribution result of the actual runoff data of the prediction period; The flood warning based on the probability distribution result includes: determining the probability that the actual runoff value at each time in the prediction period is greater than the runoff threshold of different levels at the corresponding time based on the probability distribution result; if the probability is greater than the probability threshold of the corresponding level, issuing a current level flood warning.

8. A fusion physical mechanism deep learning short and medium term runoff forecasting device, characterized in that, It includes: a data acquisition unit configured to acquire hydro-meteorological data and meteorological forecast data of a target basin; a forecast data processing unit configured to input the hydro-meteorological data and the meteorological forecast data into a trained multivariate mixed downscaling model to obtain model meteorological forecast data of a prediction period; a runoff prediction unit configured to input the model meteorological forecast data and the hydro-meteorological data into a trained mixed model to obtain forecast runoff data of the prediction period; a back-calculation prediction unit configured to determine the posterior probability density function corresponding to the forecast runoff data of the prediction period through back-calculation prediction, and determine the probability distribution result of the actual runoff data in the prediction period based on the posterior probability density function; a warning unit configured to perform flood warning based on the probability distribution result.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the method of any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 7.

Images