CN111898660A - A hydrological simulation method based on Bayesian model average fusion of multi-source data - Google Patents

Abstract

The invention discloses a hydrological simulation method based on Bayesian model average fusion of multi-source data. First, the meteorological data, hydrological series, satellite inversion precipitation and re-analysis temperature data sets of ground stations in areas with scarce data are collected; The daily bias correction method, regression correction method and equal rate correction method of digit mapping are used to establish correction models for ground observation data and simulated meteorological data sets in the same period in different months; The weight of each deviation correction scenario is optimized by the experimental probability density function, and the corrected long series of meteorological data sets are obtained; the watershed hydrological model and the long-term and short-term memory neural network model are calibrated according to the measured data, and finally the corrected long series of meteorological data sets are input to realize runoff. Process simulation. The invention can realize long-series runoff simulation in areas with scarce data, and can provide an important and highly operable reference basis for the management and planning of water resources in the basin.

Description

A hydrological simulation method based on Bayesian model average fusion of multi-source data

technical field

The invention relates to the technical field of hydrological simulation, in particular to a hydrological simulation method based on Bayesian model average fusion of multi-source data.

Background technique

High-quality long series of precipitation and temperature data are important basic data for disaster early warning and prevention, agricultural production management, ecological protection, watershed hydrological simulation, and planning and design of water conservancy projects. Traditional meteorological data mainly rely on station observations, but the density of meteorological station networks is usually small and the spatial distribution is uneven, which makes it difficult to accurately reflect the temporal and spatial variation characteristics of meteorological variables, and cannot meet the needs of engineering applications such as high-precision hydrological simulation.

In recent years, satellite telemetry technology and data inversion algorithms have developed rapidly. Precipitation quantitative observation products based on satellite remote sensing inversion have wider coverage and higher temporal and spatial resolutions, effectively making up for the shortcomings of insufficient meteorological station layout, and providing the New data references are provided for areas with no data. At the same time, with the increasing maturity of human observation methods and data assimilation technology, scholars have carried out quality control of observation data from various sources (ground, ships, radiosondes, wind balloons, aircraft, satellites, etc.), and proposed the use of numerical weather forecasting. The data assimilation technology is used to reconstruct the long-term historical climate process, the so-called reanalysis data set.

In the process of implementing the present invention, the inventor of the present application found that the method of the prior art has at least the following technical problems:

Due to the influence of remote sensing accuracy, inversion algorithm, numerical prediction model and assimilation scheme, satellite precipitation and reanalysis air temperature data have large systematic deviations, which are difficult to be directly applied to watershed hydrological simulation. Scholars at home and abroad have evaluated the applicability of the inversion dataset in the fields of meteorology, agriculture, and hydrology in different climatic regions, and a few studies have corrected the systematic bias of the precipitation temperature dataset. However, different bias correction methods have certain differences, which bring great uncertainty to the runoff simulation, and the simulation effect of the existing methods is not good.

SUMMARY OF THE INVENTION

The present invention proposes a hydrological simulation method based on Bayesian model average fusion of multi-source data, which is used to solve or at least partially solve the technical problem of poor hydrological simulation effect existing in the method in the prior art.

In order to solve the above technical problems, the present invention provides a hydrological simulation method based on Bayesian model average fusion of multi-source data, including:

S1: Collect ground observation data from ground stations in areas with scarce data. The ground observation data includes limited meteorological observation data, hydrological measured series data sets, satellite inversion precipitation data sets, and re-analysis temperature data sets;

S2: The daily deviation correction method, regression correction method and equal rate correction method based on quantile mapping are respectively used to establish a deviation correction model between the ground observation data and the simulated meteorological data set of the same period in different months. Corresponds to a bias correction model, and each bias correction model corresponds to a set of correction data sets;

S3: Using the seasonal Bayesian model averaging method, post-evaluate the three sets of correction data sets corresponding to the bias correction model established by S2, and obtain a long series of meteorological data sets;

S4: Calibrate the pre-built watershed hydrological model and long-term and short-term memory neural network model, and use the long-series meteorological data set obtained in S3 to drive the calibrated watershed hydrological model and long-term and short-term memory neural network model to output a long series of runoff process as a hydrological simulation result.

In one embodiment, S2 specifically includes:

S2.1: Use the daily deviation correction method based on quantile mapping to correct the precipitation occurrence frequency, magnitude and temperature simulation deviation month by month, and obtain the first correction data set corresponding to the deviation correction model;

S2.2: Use the regression correction method to correct the precipitation level and temperature simulation deviation month by month, and obtain the second correction data set corresponding to the deviation correction model;

S2.3: Use the equal rate correction method to correct the precipitation magnitude and temperature simulation deviation monthly, and obtain the third correction data set corresponding to the deviation correction model.

In one embodiment, S3 specifically includes:

S3.1: Construct the probability density function of meteorological correction variables according to the Bayesian full probability formula;

S3.2: Determine the corresponding weights according to the relative contribution of the bias correction effect of each bias correction model, so as to establish a seasonal Bayesian model average correction model;

S3.3: Input the long series of meteorological data sets into the seasonal Bayesian model average model established in S3.2, and obtain the corrected long series of meteorological data sets through the weighted average method of optimal weights.

In one embodiment, S4 specifically includes:

S4.1: According to the short-series runoff observation data and the meteorological observation data of the same period in the area with scarce data, construct a watershed hydrological model, and calibrate the parameters of the watershed hydrological model;

S4.2: Based on the calibrated watershed hydrological model, simulate the daily runoff process;

S4.3: Use machine learning technology to correct the simulated daily runoff process and build a long short-term memory neural network model;

S4.4: Input the long series of meteorological data sets obtained in S3 into the watershed hydrological model with a fixed rate, output the daily runoff process, and then input the output daily runoff process into the long-term and short-term memory neural network model to simulate a long series of runoff process, and take the long series of runoff processes simulated as the hydrological simulation results.

In one embodiment, S4.3 specifically includes:

Through the statistical analysis of the simulated daily runoff process and the measured daily runoff process in the area with scarce data, the lag time affecting the daily measured runoff is determined; then the long-short-term memory neural network model is used to analyze the daily runoff process simulated in step S4.2. Make corrections, where the corrected simulated runoff series is expressed as:

_Qcor (t)=F _LSTM [ _Qsim (t), _Qsim (t-1), _Qsim (t-2),…, _Qsim (tN)]

where Q _cor (t) represents the corrected runoff at time t, Q _sim (t) represents the simulated runoff of the hydrological model at time t, Q _sim (t-1) represents the simulated runoff of the hydrological model at time t-1, N represents the delay time determined by the long short-term memory neural network model; FLSTM represents the long short-term memory neural network model.

The above-mentioned one or more technical solutions in the embodiments of the present application have at least one or more of the following technical effects:

The present invention provides a hydrological simulation method based on Bayesian model average fusion of multi-source data, which gives full play to the advantages of satellite remote sensing observation data and ground meteorological station data, overcomes the low spatial resolution of observations by satellite inversion and reanalysis technology, Due to the short series of ground observation stations, a long series of meteorological data was obtained through data fusion and deviation correction technology, which improved the hydrological simulation effect, was scientific and reasonable, and was close to engineering practice; Water resources planning provides an important and operable reference basis; using a basin hydrological model with only four parameters, the daily runoff process is simulated, taking into account engineering measures such as dams, reservoirs, agricultural irrigation, water diversion and inter-basin water transfer This often leads to large errors in the basin hydrological model. Machine learning technology is used to further calibrate the simulated runoff and obtain a long series of runoff processes.

Description of drawings

In order to illustrate the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are For some embodiments of the present invention, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative efforts.

Fig. 1 is the concrete flow chart of the method of the present invention;

Figure 2 is a schematic diagram of the observation of the ground meteorological station, the grid meteorological simulation data and the runoff data;

Figure 3 is a schematic diagram of the posterior probability density function of precipitation in the constant rate correction mode.

Detailed ways

The invention provides a hydrological simulation method based on Bayesian model average fusion of multi-source data, and adopts the method based on Bayesian model average fusion of multi-source data to perform hydrological simulation on runoff in areas with scarce data, thereby improving the hydrological simulation effect.

In order to achieve the above-mentioned technical effect, the main inventive concept of the present invention is as follows: firstly collect the limited meteorological observation data, hydrological measured series, satellite-retrieved precipitation and re-analyzed temperature data sets of ground stations in areas with scarce data; The daily bias correction method, the regression correction method and the equal rate correction method are used to establish the correction model of the ground observation data and the simulated meteorological data set of the same period in different months respectively; Carry out post-evaluation, optimize the weight of each bias correction scenario through the posterior probability density function, and obtain a long series of corrected meteorological data sets; calibrate the basin hydrological model and long-term short-term memory neural network model according to the measured data, and finally input the corrected long-term data set. A series of meteorological data sets to realize runoff simulation in areas with scarce data.

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

A hydrological simulation method based on Bayesian model average fusion of multi-source data according to an embodiment of the present invention, firstly collects limited meteorological observation data, hydrological measured series, satellite inversion of precipitation and re-analyzed temperature data sets of ground stations in areas with scarce data; Then, the daily deviation correction method, regression correction method and equal rate correction method based on quantile mapping are used to establish three sets of ground observation data in different months and the correction model of the simulated meteorological data set in the same period respectively; The model averaging method is used to perform post-evaluation on three sets of correction data sets, and optimize the weight of each deviation correction scheme through the posterior probability density function; calibrate the watershed hydrological model and the long-term and short-term memory neural network model according to the measured data, and finally input the corrected long-term and short-term memory neural network model. A series of meteorological data sets to achieve runoff simulation in areas with scarce data. The specific process is shown in Figure 1.

The technical solutions of the present invention will be further described in detail below through the examples and in conjunction with the accompanying drawings. The method of the present invention specifically includes:

S1: Collect ground observation data from ground stations in areas with scarce data. The ground observation data includes limited meteorological observation data, hydrological measured series data sets, satellite inversion precipitation data sets, and reanalysis temperature data sets.

In the specific implementation, the satellite retrieval precipitation product used in this embodiment is the MSWEP-V2 data set. This product integrates multiple satellite retrieval data sources, 76747 global ground station observation data and reanalysis data sources, and also uses global 13762 data sources. Based on the water balance, the measured data of the two runoff stations have corrected the deviation of the terrestrial precipitation process, which is one of the precipitation data sources with the highest temporal and spatial accuracy in the world.

Further, the reanalysis temperature data set used in this embodiment is the fifth-generation reanalysis climate product ERA5 of the European Center for Medium-Range Weather Forecasts; the horizontal resolution of the hourly analysis field of this data set is 31km, the vertical layer is 137 layers, and the top layer is 137 layers. Reach a height of 0.01hPa; ERA5 adopts the Cycle31r2 model version of the integrated forecasting system. Based on the spectral harmonic resolution T255, the simplified Gaussian grid (N128) data is interpolated to 0.25° to 2.5° through bilinear interpolation technology. A variety of different resolution rasters are currently one of the global reanalysis data with the highest spatial and temporal resolution.

As shown in Figure 2, a schematic diagram of the meteorological observation data, runoff data and grid simulation meteorological data at the ground station is given. For the watershed of the embodiment, the length of the ground observation meteorological data and the runoff data series is short, so it is necessary to collect more data. The long-term meteorological simulation data can only be realized by driving the hydrological model to obtain the corrected long series of meteorological data sets through the deviation correction model.

S2: The daily deviation correction method, regression correction method and equal rate correction method based on quantile mapping are respectively used to establish a deviation correction model between the ground observation data and the simulated meteorological data set of the same period in different months. Corresponds to a bias correction model, and each bias correction model corresponds to a set of correction data sets.

In one embodiment, S2 specifically includes:

S2.1: Use the daily deviation correction method based on quantile mapping to correct the precipitation occurrence frequency, magnitude and temperature simulation deviation month by month, and obtain the first correction data set corresponding to the deviation correction model;

Specifically, the precipitation threshold of each month in the simulated series is determined according to the occurrence frequency of the measured daily precipitation in different months. When the daily precipitation is higher than this threshold, it is judged as a rainy day, otherwise it is no rainy day; then the measured daily precipitation (temperature) is calculated. and the systematic deviation of the frequency distribution function of the historical simulation series in each month, then calculate the correction coefficient corresponding to each quantile, and finally use the coefficient to correct the simulated long series of meteorological data:

和

分别为校正后第m月的日降水和气温系列；P_G,m和T_G,m分别为历史期校正前第m月的日降水和气温系列；F_obsP,m和F_GP,m(F_obsT,m和F_GT,m) 分别为历史期日降水(气温)实测和模拟系列的累积分布函数。where:

and

are the daily precipitation and temperature series of the mth month after correction; P _G,m and T _G,m are the daily precipitation and temperature series of the mth month before the correction in the historical period, respectively; F _obsP,m and F _GP,m (F _obsT,m and F _GT,m ) are the cumulative distribution functions of the observed and simulated series of daily precipitation (temperature) in the historical period, respectively.

S2.2: Use the regression correction method to correct the precipitation level and temperature simulation deviation month by month, and obtain the second correction data set corresponding to the deviation correction model;

Specifically, a regression relationship is established (that is, 12 regression correction models are established) for the measured daily precipitation (air temperature) and the satellite fusion daily precipitation (re-analysis data set air temperature) of the meteorological stations in each month, as follows:

P _obs,m =a _P,m +b _P,m · _PG,m +ε

T _obs,m =a _T,m +b _T,m ·T _G,m +ε (2)

In the formula: P _obs,m and T _obs,m are the measured daily precipitation and daily temperature of the mth month, respectively, P _G,m and T _G,m are the simulated precipitation and temperature series in the same period; a _P,m and b _P,m (a _T,m and b _T,m ) are the regression coefficients of the precipitation (air temperature) series, respectively, and ε represents the model residual.

S2.3: Use the equal rate correction method to correct the precipitation magnitude and temperature simulation deviation monthly, and obtain the third correction data set corresponding to the deviation correction model.

Specifically, the equal-rate correction method assumes that the monthly deviations of the satellite-retrieved precipitation (or re-analyzed temperature data) and the ground observation series are consistent in different periods. First, the correction factor for each month is calculated based on the ground observation information, and then the factor Applied to a long series of simulated datasets in the same month. This method is easy to operate and has good effect, and has been widely used in the field of satellite retrieval precipitation product correction in recent years. For each month of a meteorological station, this paper calculates the deviation ratio (precipitation) or absolute deviation (air temperature) of the month based on the observation data and simulation series, and then uses the following formula to correct the satellite precipitation and re-analyze the temperature series after obtaining the corresponding correction factor:

In the formula: N represents the total number of observation days in the mth month of the station, and i represents the daily precipitation or temperature series.

S3: Using the seasonal Bayesian model averaging method, post-evaluate the three sets of correction data sets corresponding to the bias correction model established by S2, and obtain a long series of meteorological data sets.

Among them, S3 specifically includes:

S3.1: Construct the probability density function of meteorological correction variables according to the Bayesian full probability formula;

Specifically, let S be the correction variable, R=[D, O] represent the input data of the model (where D is the correction series of each method in the training period, and O is the measured series), f=[f ₁ , f ₂ ,..., f _K ] is the output result of K different correction modes, and the probability density function of S obtained by the Bayesian total probability formula is as follows:

In the formula: p _k (S|f _k , R) is the probability density function of the correction value S of the Kth correction mode f _k under the condition of given data R; p(f _k | R) is the given training data R The posterior probability density function for the kth corrected mode.

As shown in Fig. 3, a schematic diagram of the posterior probability density function of precipitation in the equal-rate correction mode is given.

S3.2: Determine the corresponding weights according to the relative contribution of the bias correction effect of each bias correction model, so as to establish a seasonal Bayesian model average correction model;

Specifically, firstly, the observation series of meteorological stations and the simulation series obtained by each correction method are normally transformed by the Box-Cox function, and then the estimation results of various models are weighted and averaged based on the assumption of normal linear distribution:

表示均值为f_k，方差为

的正态分布；E表示函数期望值，w_k为第k个偏差校正模式的权重。where:

means that the mean is f _k and the variance is

The normal distribution of ; E represents the expected value of the function, and w _k is the weight of the k-th bias correction mode.

Further, this embodiment takes K=3.

S3.3: Input the long series of meteorological data sets into the seasonal Bayesian model average model established in S3.2, and obtain the corrected long series of meteorological data sets through the weighted average method of optimal weights.

S4: Calibrate the pre-built watershed hydrological model and long-term and short-term memory neural network model, and use the long-series meteorological data set obtained in S3 to drive the calibrated watershed hydrological model and long-term and short-term memory neural network model to output a long series of runoff process as a hydrological simulation result.

In one embodiment, S4 specifically includes:

S4.1: According to the short-series runoff observation data and the meteorological observation data of the same period in the area with scarce data, construct a watershed hydrological model, and calibrate the parameters of the watershed hydrological model;

Specifically, the basin hydrological model is the GR4J hydrological model, which is a lumped conceptual hydrological model with only 4 parameters. This model has the characteristics of simple structure, few parameters and high precision, and has been widely used. The model is mainly composed of two nonlinear reservoirs, namely the runoff reservoir and the confluence reservoir. Calibration is calibration.

S4.2: Based on the calibrated watershed hydrological model, simulate the daily runoff process;

Specifically, the simulated runoff series is expressed as:

Q _sim = F _GR4J [Prep, Tmean, Latitude, BasinArea, ParameterX] (6)

In the formula: Qsim represents the simulated runoff series, Prep represents the rainfall series after downscaling, Tmean represents the daily average temperature series after downscaling, Latitude represents the mean latitude of the watershed, BasinArea represents the area of the watershed, ParameterX represents the rate in step 4.1 Determine the obtained model parameters, FGR4J represents the GR4J model.

S4.3: Use machine learning technology to correct the simulated daily runoff process and build a long short-term memory neural network model;

Among them, S4.3 specifically includes:

Through the statistical analysis of the simulated daily runoff process and the measured daily runoff process in the area with scarce data, the lag time affecting the daily measured runoff is determined; then the long-short-term memory neural network model is used to analyze the daily runoff process simulated in step S4.2. Make corrections, where the corrected simulated runoff series is expressed as:

_Qcor (t)=F _LSTM [ _Qsim (t), _Qsim (t-1), _Qsim (t-2),…, _Qsim (tN)] (7)

where Q _cor (t) represents the corrected runoff at time t, Q _sim (t) represents the simulated runoff of the hydrological model at time t, Q _sim (t-1) represents the simulated runoff of the hydrological model at time t-1, N represents the delay time determined by the long short-term memory neural network model; FLSTM represents the long short-term memory neural network model.

Further, the LSTM model is trained using the mini-batch gradient descent method, which is a conventional technique in the art.

S4.4: Input the long series of meteorological data sets obtained in S3 into the watershed hydrological model with a fixed rate, output the daily runoff process, and then input the output daily runoff process into the long-term and short-term memory neural network model to simulate a long series of runoff process, and take the long series of runoff processes simulated as the hydrological simulation results.

The specific embodiments described in the present invention are merely illustrative of the spirit of the present invention. Those skilled in the art to which the present invention pertains can make various modifications or additions to the described specific embodiments or substitute in similar manners, but will not deviate from the spirit of the present invention or go beyond the scope defined by the appended claims .

Claims (5)

1. A hydrologic simulation method for averagely fusing multi-source data based on a Bayesian mode is characterized by comprising the following steps of:

s1: collecting ground observation data of ground stations in a scarce data area, wherein the ground observation data comprises limited meteorological observation data, a hydrologic actual measurement series data set, a satellite inversion precipitation data set and a re-analysis gas temperature data set;

s2: respectively adopting a daily deviation correction method, a regression correction method and an equal rate correction method based on quantile mapping to establish deviation correction models of ground observation data and a synchronous simulated meteorological data set in different months, wherein each correction method corresponds to one deviation correction model, and each deviation correction model corresponds to one set of correction data set;

s3: carrying out post-evaluation on the three sets of correction data sets corresponding to the deviation correction model established in the step S2 by adopting a seasonal Bayesian mode averaging method to obtain a long-series meteorological data set;

s4: and (4) calibrating the pre-constructed watershed hydrological model and the long-short term memory neural network model, driving the calibrated watershed hydrological model and the long-short term memory neural network model by using the long-series meteorological data set obtained in S3, outputting the long-series runoff process, and taking the long-series runoff process as a hydrological simulation result.

2. The hydrological simulation method of claim 1, wherein S2 specifically comprises:

s2.1: correcting precipitation occurrence frequency, magnitude and air temperature simulation deviation month by adopting a daily deviation correction method based on quantile mapping to obtain a first correction data set corresponding to the deviation correction model;

s2.2: correcting the precipitation magnitude and the air temperature simulation deviation month by adopting a regression correction method to obtain a second correction data set corresponding to the deviation correction model;

s2.3: and correcting the precipitation magnitude and the air temperature simulation deviation month by adopting an equal rate correction method to obtain a third correction data set corresponding to the deviation correction model.

3. The hydrological simulation method of claim 1, wherein S3 specifically comprises:

s3.1: constructing a probability density function of a meteorological correction variable according to a Bayes total probability formula;

s3.2: determining corresponding weights according to the relative contribution of the deviation correction effect of each deviation correction model, thereby establishing a seasonal Bayesian mode average correction model;

s3.3: and inputting the long-series meteorological data sets into a seasonal Bayesian mode average model established in S3.2, and obtaining the corrected long-series meteorological data sets by a weighted average method of optimal weight.

4. The hydrological simulation method of claim 1, wherein S4 specifically comprises:

s4.1: constructing a basin hydrological model according to short series runoff observation data of scarce data areas and meteorological observation data of the same period, and calibrating parameters of the basin hydrological model;

s4.2: simulating to obtain a daily runoff process based on a calibrated watershed hydrological model;

s4.3: correcting the daily runoff process obtained by simulation by adopting a machine learning technology, and constructing a long-term and short-term memory neural network model;

s4.4: inputting the long series meteorological data set obtained in the S3 into a well-calibrated watershed hydrological model, outputting a daily runoff process, inputting the output daily runoff process into the long and short term memory neural network model, simulating the long series runoff process, and taking the simulated long series runoff process as a hydrological simulation result.

5. The hydrological simulation method of claim 1, wherein S4.3 specifically comprises:

determining the time delay influencing the daily actual measurement runoff by carrying out statistical analysis on the simulated daily runoff process and the actual measurement daily runoff process in the scarce data area; and correcting the daily runoff process simulated in the step S4.2 by adopting a long-short term memory neural network model, wherein the corrected simulated runoff series is represented as:

Q_cor(t)＝F_LSTM[Q_sim(t),Q_sim(t-1),Q_sim(t-2),···,Q_sim(t-N)]

in the formula: q_cor(t) represents the corrected runoff at time t, Q_sim(t) simulated runoff of the hydrological model at time t, Q_sim(t-1) representing simulated runoff of the hydrological model at the t-1 moment, and N representing the time lag determined by the long-term and short-term memory neural network model; FLSTM represents a long-short term memory neural network model.

Images