Artificial intelligence-based flood probability forecasting method and device

Abstract

The invention belongs to the technical field of data processing, and particularly provides a flood probability forecasting method and device based on artificial intelligence, and the method comprises the steps: obtaining downscaled long-series meteorological observation data through employing a statistical downscaling model based on ground observation data and satellite remote sensing inversion data; constructing a deep learning model considering the hydrological process; reconstructing to obtain a long-series daily runoff process of the scarcity data drainage basin, and extracting a flood event by adopting an over-quantitative sampling method; adopting a plurality of artificial intelligence models suitable for flood magnitude simulation to construct a flood simulation model; based on the simulation result and the actually measured flood magnitude, a Bayesian mode averaging method is adopted to deduce the weight of each model, and an artificial intelligence-Bayesian coupling model is constructed; a Copula function is combined to construct a distribution function of actually measured flow and forecast flow, and an artificial intelligence-Bayesian coupling model is adopted to carry out flood forecasting. The method is used for solving the defect that the uncertainty and the artificial intelligence technology are not fully considered for flood forecasting.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and in particular, relates to a flood probability forecasting method and device based on artificial intelligence. Background Art

[0002] In recent years, extreme precipitation and floods have occurred frequently, posing a serious threat to the ecological, flood control, energy, and food security of both countries. To address these challenges, there is an urgent need to study the mechanisms underlying extreme precipitation and flood events and improve forecast accuracy to support disaster prevention and mitigation efforts. Domestic and international researchers have developed flood forecasting methods by establishing linear or nonlinear relationships between meteorological variables and floods. This approach requires a long series of meteorological and hydrological data. However, hydrological and meteorological data are extremely scarce in most regions of my country, with some areas having only a limited amount of measured hydrological and meteorological data. Therefore, effective flood forecasting in basins with limited data presents a major challenge for hydrologists. Satellite remote sensing primarily uses sensors onboard meteorological satellites to measure meteorological data. However, satellite remote sensing technology suffers from large grid sizes and low spatial resolution, making it difficult to directly meet practical requirements. Therefore, the rational use of satellite remote sensing data is crucial for flood forecasting in data-scarce basins. At the same time, hydrological models are suitable for simulating runoff processes under natural conditions. Engineering measures such as dams, reservoirs, agricultural irrigation, water diversion and inter-basin water transfer often destroy the consistency of the underlying surface, resulting in large errors in the basin hydrological model and restricting the accuracy of hydrological simulation.

[0003] Inputting meteorological forecast data into hydrological models for flood forecasting is a common approach. However, deterministic forecasts generated by a single model often have high uncertainty, making it difficult to obtain objective and accurate risk assessments for flood control decisions, such as reservoir operation. Integrating the forecast results of multiple hydrological models and employing statistical post-processing methods to construct a hydrological ensemble forecasting method is an important approach to quantifying or reducing hydrological model uncertainty. By quantifying the uncertainty of various hydrological model structures and parameters, not only can deterministic forecast values be derived, but forecast uncertainty information can also be provided in the form of quantiles, confidence intervals, or density functions. This is more scientific and reasonable than deterministic forecasts from a single model, helping decision makers quantitatively identify forecast risks. Overall, existing literature rarely considers the errors caused by human interference in runoff simulations, fails to address the challenge of long-series flood forecasting in basins with scarce data, and the application of artificial intelligence technology to probabilistic flood forecasting remains under discussion. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a flood probability forecasting method and device based on artificial intelligence, so as to solve the defects of the prior art in flood forecasting that fail to fully consider uncertainty and artificial intelligence technology.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is: a flood probability forecasting method based on artificial intelligence, comprising the following steps: Step 1: Based on ground observation data and satellite remote sensing inversion data, a statistical downscaling model is constructed using the equal-rate correction method. The downscaled long-series meteorological observation data, including long-series precipitation and temperature observation data, are obtained through the statistical downscaling model. Step 2: Based on the short-series runoff observation data and the downscaled long-series meteorological observation data of the data-scarce basin, a basin hydrological model is established. The long-short-term memory model is used for calibration to construct a deep learning model that takes into account the hydrological process. Step 3: Input the long series of meteorological observation data obtained in step 1 into the deep learning model considering the hydrological process established in step 2, reconstruct the long series of daily runoff processes in the basin with scarce data, and extract flood events using the super-quantitative sampling method; Step 4: Based on the meteorological data and the extracted flood events, K artificial intelligence models suitable for flood magnitude simulation are used to construct a flood simulation model, and the model parameters are calibrated; Step 5: Based on the simulation results of the K flood simulation models and the measured flood magnitude, the Bayesian pattern averaging method is used to deduce the weight of each model, thereby constructing an artificial intelligence-Bayesian coupling model; Step 6: Combine the Copula function to construct the distribution function of the measured flow and the forecast flow, and use the artificial intelligence-Bayesian coupling model to carry out flood forecasting.

[0006] In a preferred embodiment, step 1 includes the following sub-steps: S1.1. Collect ground observation data from ground stations in data-scarce areas. Ground observation data include limited meteorological observation data and hydrological observation series datasets. Satellite remote sensing inversion data include satellite-derived precipitation products and reanalysis datasets. S1.2. Use the equal-rate correction method to construct a statistical downscaling model and correct the precipitation magnitude and temperature simulation bias on a monthly basis; The isorelative correction method assumes that the monthly deviations between satellite-retrieved precipitation and reanalysis temperature and ground-based observations are consistent over time. First, a correction factor is calculated for each month based on ground-based observations, and then the factor is applied to the long series of simulated data sets for the same month. The precipitation deviation ratio and absolute temperature deviation for that month were calculated based on the observation data and satellite remote sensing inversion data, respectively. After obtaining the corresponding correction factors, the satellite inverted precipitation series and reanalysis temperature series were corrected using the following formula: (1); Where: N Indicates site m Total number of observation days per month;i Represents a daily precipitation or temperature series; Indicates that the satellite-retrieved precipitation m The original data series for the month, Indicates the corrected satellite-retrieved precipitation in the m Monthly series; Indicates that the satellite-retrieved precipitation m Month i The original data of the day, Indicates that the weather station is m Month i Daily observed precipitation data; Reanalysis temperature in the m The original data series for the month, The corrected reanalysis temperature product is m Monthly series; Reanalysis temperature product in the m Month i The original data of the day, Indicates that the weather station is m Month i Observed temperature data for the day.

[0007] In a preferred embodiment, step 2 includes the following sub-steps: S2.1. For basins with scarce data, the Xin'an River model is constructed using short-term runoff observation data and long-term meteorological observation data from the same period after statistical downscaling, and the model parameters are calibrated. S2.2. Based on the calibrated Xin'anjiang model, the first simulated daily runoff process is simulated. By statistically analyzing the first simulated daily runoff process and the measured daily runoff process, the lag time that affects the measured daily runoff is determined. N ; S2.3. Use the long series of precipitation and temperature observations corrected in step 1 to drive the Xin'anjiang model calibrated in step 2.1 to obtain a long series of daily runoff simulations, and divide the simulations into three types of daily runoff series with different magnitudes. S2.4. Based on the three types of daily runoff series with different magnitudes, a long-short-term memory neural network model is constructed to correct the first simulated daily runoff process. After correction, the second simulated runoff series is obtained, which is expressed as: ](2); Where, express t The second simulated runoff after time correction, Indicates that the Xinanjiang model is t The first simulated daily runoff at time, Indicates that the Xinanjiang model is t-The first simulated daily runoff at time 1, N represents the lag time determined in step 2.2; F LSTM Represents an LSTM model.

[0008] In the preferred embodiment, in step S2.3, the method for dividing the simulation series into three types of daily runoff series of different magnitudes is: calculating the 35% quantile and 65% quantile of the daily runoff series, and dividing the daily runoff series into a first type below the 35% quantile, a second type between the 35% quantile and the 65% quantile, and a third type exceeding the 65% quantile.

[0009] In the preferred scheme, in the step 3, a long series of daily runoff processes in the basin with scarce data is reconstructed, and the 80% quantile of the daily runoff is used as the threshold, and the daily runoff exceeding the threshold is identified and used as the flood event magnitude; for daily runoff that continuously exceeds the threshold, the average of the daily runoff during this period is taken as the flood event magnitude; when the calculated flood magnitude exceeds the 95% quantile of the daily runoff, the flood event is identified as an outlier, and the 95% quantile of the daily runoff is taken as the flood magnitude of the event.

[0010] In a preferred embodiment, step 4 includes the following sub-steps: S4.1. Based on satellite remote sensing inversion data, collect hourly meteorological and hydrological observation data, including 2-meter air temperature, air pressure, dew point temperature, snowfall, precipitation, wind speed, shortwave radiation, and longwave radiation. Obtain daily average 2-meter air temperature, daily average dew point temperature, daily snowfall, daily precipitation, daily average air pressure, daily wind speed, daily shortwave radiation, and daily longwave radiation through time scale conversion. S4.2. Calculate specific humidity and relative humidity based on the Clausius-Clapeyron thermodynamic equation; The Clausius-Clapeyron thermodynamic equation is defined as follows: (3); Where: is the first integration constant, which is 273.16 K; is the second integral constant, which is 611 Pa; is the latent heat of vaporization constant, take ; is the water vapor gas constant, take ; T is the air temperature, which is the input variable of the Clausius-Clapeyron thermodynamic equation; The 2-meter air temperature and dew point temperature of the meteorological and hydrological observation samples of each grid point in the study basin are substituted as input variables into the Clausius-Clapeyron thermodynamic equation to calculate the relative humidity of each grid point. RH , , where T dew is the dew point temperature, T 2m The air temperature at 2m; Specific humidity q is the ratio of water vapor mass to total air mass mass, using the ground pressure in the reanalysis data p And dew point temperature, the formula is as follows: (4); S4.3. Use the Thiessen polygon method to derive average meteorological variables for basins with scarce data. Combine the Honghu flood events in the basin with the basin's meteorological variables. The meteorological variables include the average daily temperature, specific humidity, relative humidity, snowfall, precipitation, wind speed, shortwave radiation, and longwave radiation on the day of the flood event and 1-3 days before. Use a random forest model to screen key meteorological factors. S4.4. Based on the reanalysis dataset, using the key meteorological factors selected in step S4.3 as input and flood magnitude as the target variable, construct K artificial intelligence models suitable for flood magnitude simulation.

[0011] In the preferred scheme, in step S4.4, the K artificial intelligence models suitable for flood magnitude simulation include: linear regression, logistic regression, decision tree, random forest, support vector machine, K nearest neighbor algorithm, naive Bayes classifier, neural network, convolutional neural network, recurrent neural network, long short-term memory network, generative adversarial network, Transformer, XGBoost and Markov decision process model.

[0012] In a preferred embodiment, step 5 includes the following sub-steps: S5.1. Use meteorological data from the reanalysis dataset of satellite remote sensing inversion data as input to drive the K artificial intelligence models suitable for flood magnitude simulation constructed in step 4 to obtain a simulated magnitude series of flood events in the data-scarce basin. S5.2. Based on the simulated flood magnitude results and measured flood indices of K AI models suitable for flood magnitude simulation, a probability density function of the simulated flood magnitude is constructed according to the Bayesian total probability formula. Let S be the simulated flood magnitude, and R = [D, O] represent the model input data, where D is the simulation series of each AI model during the training period, and O is the measured series. The output results of K different artificial intelligence models are obtained by the Bayesian total probability formula. The probability density function of S is as follows: (5); Where: For the k An artificial intelligence model for flood magnitude simulation The probability density function of the simulated value S under the given data R; For a given training data R k The posterior probability density function of the artificial intelligence models suitable for flood magnitude simulation, K represents the number of artificial intelligence models; S5.3. Determine the corresponding weights based on the relative contributions of each AI model to simulate flood magnitude, thereby establishing a Bayesian pattern average correction model. The specific operations are as follows: First, the flood magnitude observation series and the simulation series obtained by each artificial intelligence model are transformed into normality using the Box-Cox function. Then, the estimation results of multiple models are weighted averaged based on the normal linear distribution assumption: (6); Where: Indicates the mean , the variance is Normal distribution; E represents the expected value of the function, For the k The weights of an AI model for flood magnitude simulation; S5.4. Input the downscaled long series meteorological data set into the K artificial intelligence models suitable for flood magnitude simulation constructed in step 4 respectively, and then input the simulation results of the K artificial intelligence models suitable for flood magnitude simulation into the Bayesian pattern averaging model, and obtain high-precision flood simulation results through the weighted averaging method of preferred weights.

[0013] In a preferred embodiment, step 6 includes the following steps: S6.1. Obtain weather forecast information for watersheds where data is scarce, including daily average temperature, specific humidity, relative humidity, snowfall, precipitation, wind speed, shortwave radiation, and longwave radiation for the next 1-7 days; S6.2. Use the Thiessen polygon method to calculate the regional average weather forecast information for the watershed where data are scarce; S6.3. Obtain the reanalysis meteorological data for the past three days from the reanalysis dataset and use the Thiessen polygon method to obtain the regional average reanalysis meteorological data for the data-scarce basin. S6.4. Use the regional average weather forecast for the next day for the data-scarce basin obtained in step 6.2 and the regional average reanalysis weather data for the past 1 to 3 days obtained in step 6.3 as input to drive the artificial intelligence-Bayesian coupling model established in step 5 to obtain the flood forecast for the next day. S6.5, using steps S6.1 to S6.4, using weather forecast information with a longer future horizon as input, thereby obtaining flood forecast values for the next 2-3 days; S6.6. Use the retrospective forecasting method to construct the joint distribution function of flood observations and flood forecasts over the historical period: make H k 、 S k Respectively represent k The actual flood magnitude at each moment and the deterministic flood forecast magnitude, h k 、 s k Then H k 、 S k The realized value of and They are H k 、 S k The marginal distribution function of and The corresponding H k 、 S k The probability density function of , according to Sklar's theorem, there exists a n -Copula function C , so that the following formula holds: (7); Where: Copula function parameters; is the density function of the two-dimensional Copula function; The joint distribution function representing the actual flood magnitude and the predicted magnitude; S6.7. Derivation of the posterior distribution function of flood magnitude based on the Copula function and probabilistic forecasting; Given forecast value s k ,exist S k = s k hour, H k The conditional distribution function of for: (8); H k The density function is expressed as: (9); and When the deterministic forecast flood magnitude information is given, the measured flood magnitude h k The posterior probability distribution function and posterior probability density function are used to describe the uncertainty of the forecast; the expected value of the posterior probability density function is taken as the post-processing deterministic forecast result, and the confidence level is obtained at the same time The probability forecast interval of the flood magnitude under the condition of , and the expected value of the posterior probability density function of the measured flood magnitude h ke Solve the following equation: (13).

[0014] The present invention also provides an artificial intelligence-based flood probability forecasting device for use in the above-mentioned flood probability forecasting method, comprising: The first determination module is used to construct a statistical downscaling model based on ground observation data and satellite remote sensing inversion data using the equal-rate correction method to obtain long-series precipitation and temperature observation data; An analysis module is used to build a watershed hydrological model based on short-series runoff observations and downscaled meteorological data in data-scarce watersheds; A module was built to use a random forest model to select key physical factors affecting runoff, which were further corrected using a long-short-term memory model, thereby constructing a deep learning model that considers hydrological processes. The second determination module inputs the long series of meteorological observation data obtained by downscaling into a deep learning model that considers hydrological processes, reconstructs the long series of daily runoff processes in the basin with scarce data, extracts flood events using the ultra-quantitative sampling method, and constructs an artificial intelligence-Bayesian coupled model; The prediction module combines the Copula function to construct the distribution function of measured flow and forecast flow, and uses the artificial intelligence-Bayesian coupling model to carry out flood forecasting.

[0015] The present invention provides an artificial intelligence-based flood probability forecasting method and device, which has the following beneficial effects: 1. By coupling the physical mechanism of flood formation, the uncertainty of hydrological processes, artificial intelligence and satellite remote sensing, it provides an important and highly operational reference for flood forecasting in basins with scarce data, and provides engineering reference value for responding to climate disasters, disaster prevention and mitigation, and integrated water resources management.

[0016] 2. The satellite remote sensing inversion data (such as MSWEP-V2 precipitation products and ERA5 reanalysis temperature) are corrected for monthly bias through the equal-rate correction method, which solves the problems of large grid scale and insufficient accuracy of satellite data, generates high-precision long-series meteorological data, provides reliable basic input for data-scarce basins, and fills the temporal and spatial gaps in ground observation data.

[0017] 3. A deep learning model was constructed by combining the Xin'anjiang model with the LSTM neural network. Downscaled meteorological data was used to reconstruct the long series of daily runoff processes, and flood events were accurately identified through ultra-quantitative sampling. This solved the simulation error problem of traditional hydrological models under the interference of human activities, and improved the integrity of runoff data and the recognition of flood events.

[0018] 4. Build a flood simulation model library containing 15 AI models. Use Bayesian mode averaging to dynamically assign weights based on model performance, forming an AI-Bayesian coupled model. This avoids the one-sidedness of a single model, reduces model structural uncertainty through weighted averaging, and significantly improves flood simulation accuracy. The probabilistic framework of BMA provides a reliable foundation for subsequent probabilistic forecasts.

[0019] 5. Probabilistic forecasting based on Copula functions: This method constructs a joint distribution function of flood observations and forecasts through backcasting. Using the Copula function, it infers the posterior distribution of flood magnitude and provides probabilistic forecast results with confidence intervals (e.g., 90% probability intervals). Compared to traditional deterministic forecasting, this method quantitatively describes forecast uncertainty, provides a risk quantification basis for reservoir operation and flood control decisions, and supports decision makers in developing more scientific emergency response plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1 A schematic diagram of the method flow provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the data of ground meteorological station observations, satellite remote sensing data and runoff data; Figure 3 Schematic diagram of the change in correlation coefficient between daily measured runoff and simulated runoff at different lag times; Figure 4 Schematic diagram of the probabilistic flow forecast process and 90% confidence level forecast interval provided by the present invention; Figure 5 A block diagram of a flood probability forecasting device provided by an exemplary embodiment is shown. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0023] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0024] Example 1: like Figure 1 As shown, a flood probability forecasting method based on artificial intelligence includes the following steps: Step 1: Based on ground observation data and satellite remote sensing inversion data, a statistical downscaling model is constructed using the equal-rate correction method. The downscaled long series of meteorological observation data, including long series of precipitation and temperature observation data, are obtained through the statistical downscaling model.

[0025] like Figure 2 As shown in FIG, a schematic diagram of ground meteorological station observations, satellite remote sensing data and runoff data is given. For the watershed in the embodiment, there is a mismatch between ground meteorological data and runoff data, and they cannot be used directly to calibrate the hydrological model. Therefore, it is necessary to use relatively complete satellite telemetry data and statistical downscaling models to obtain a long series of corrected satellite telemetry data, so as to establish a hydrological model with the scarce runoff observation series and finally realize the watershed hydrological simulation.

[0026] Step 1 includes the following sub-steps: S1.1. Collect ground observation data from ground stations in areas with scarce data. Ground observation data include limited meteorological observation data and hydrological observation series data sets. Satellite remote sensing inversion data include satellite inversion precipitation products and reanalysis data sets.

[0027] In the specific implementation, the satellite-derived precipitation product used in this embodiment is the MSWEP-V2 dataset, which integrates multiple satellite inversion data sources, 76,747 global ground station observation data and reanalysis data sources. It also uses measured data from 13,762 runoff stations around the world to perform bias correction on the land precipitation process based on water balance. It is currently one of the precipitation data sources with the highest temporal and spatial accuracy in the world.

[0028] Furthermore, the reanalysis temperature dataset used in this embodiment is the fifth-generation reanalysis climate product ERA5 of the European Centre for Medium-Range Weather Forecasts. The horizontal resolution of the hourly analysis field of this dataset is 31 km, and the vertical stratification is 137 layers, with the top layer reaching an altitude of 0.01 hPa. ERA5 uses the Cycle31r2 model version of the integrated forecast system. Based on the spectral harmonic resolution T255, it interpolates the simplified Gaussian grid (N128) data to various resolution grids ranging from 0.25° to 2.5° through bilinear interpolation technology. It is one of the global reanalysis data with the highest temporal and spatial resolutions currently available.

[0029] Furthermore, this embodiment collects the daily runoff observation data of the outlet section of the watershed where the data is scarce and the observation data of the ground meteorological station.

[0030] S1.2. Use the equal-rate correction method to construct a statistical downscaling model to correct the precipitation magnitude and temperature simulation bias on a monthly basis.

[0031] The isorelative correction method assumes that the monthly deviations between satellite-retrieved precipitation and reanalysis temperature and ground-based observations are consistent over time. First, a correction factor is calculated for each month based on ground-based observations, and then the factor is applied to the long series of simulated data sets for the same month. In this embodiment, for each month at the meteorological station, the precipitation deviation ratio or absolute temperature deviation of that month is calculated based on the observation data and satellite remote sensing inversion data. After obtaining the corresponding correction factors, the satellite inverted precipitation series and reanalysis temperature series are corrected using the following formula: (1); Where: N represents the total number of observation days at the station in month m, i represents the daily precipitation or temperature series, Indicates that the satellite-retrieved precipitation m The original data series for the month, Indicates the corrected satellite-retrieved precipitation in the m Monthly series; Indicates that the satellite-retrieved precipitation m Month i The original data of the day, Indicates that the weather station is m Monthi Daily observed precipitation data; Reanalysis temperature in the m The original data series for the month, The corrected reanalysis temperature product is m Monthly series; Reanalysis temperature product in the m Month i The original data of the day, Indicates that the weather station is m Month i Observed temperature data for the day.

[0032] Step 2: Based on the short-term runoff observation data and downscaled long-term meteorological observation data of the data-scarce basin, a basin hydrological model is established. The model is calibrated using a long-short-term memory model to construct a deep learning model that takes into account the hydrological process. This includes the following sub-steps: S2.1. For basins with scarce data, the Xin'anjiang model is constructed using short-series runoff observation data and long-series meteorological observation data after statistical downscaling over the same period, and the model parameters are calibrated.

[0033] The Xin'anjiang Model, a hydrological model developed by Professor Zhao Renjun's team at the East China Institute of Water Resources (now Hohai University), is a rare Chinese hydrological model with global influence. It is a lumped hydrological model (or a decentralized hydrological model when divided into subbasins) and can be used during the wet season in humid and semi-humid regions. The Xin'anjiang Model uses a lumped model for small basins and a decentralized model for larger basins. It divides the entire basin into numerous unit basins, calculating runoff generation and confluence for each unit basin to determine the outlet flow process. Flood routing is then performed for the river channel below the outlet to determine the flow process at the basin outlet. By summing the outflow processes for each unit basin, the total outflow process for the basin is calculated.

[0034] This model calculates basin evapotranspiration using a three-layer evapotranspiration model, calculates total runoff generated by rainfall based on the concept of full storage runoff, and uses basin storage curves to account for the impact of underlying surface unevenness on variations in runoff area. Regarding runoff composition, for the three-water source scenario, the runoff is divided into saturated surface runoff, subsurface runoff, and groundwater runoff using a free-water reservoir with a finite volume, metering holes, and bottom holes, based on the "hillslope hydrology" runoff theory. Confluence calculations generally use the unit line method for surface runoff confluence per unit area, while the linear reservoir method is used for the confluence of subsurface runoff and groundwater runoff.

[0035] Calibration of the Xin'anjiang model is a routine technique in this field.

[0036] S2.2. Based on the calibrated Xin'anjiang model, the first simulated daily runoff process is simulated. By statistically analyzing the first simulated daily runoff process and the measured daily runoff process, the lag time that affects the measured daily runoff is determined. N .

[0037] like Figure 3 As shown in the figure, a schematic diagram of the change in the correlation coefficient between the daily measured runoff and the first simulated runoff under different lag times is given; the correlation coefficient between the first simulated runoff and the measured runoff generally decreases gradually with the extension of the lag time; further, a suitable correlation threshold is selected to determine the length of simulated runoff for establishing an artificial intelligence model with the measured runoff; for example, 0.5 can be taken.

[0038] S2.3. Use the long series of precipitation and temperature observations corrected in step 1 to drive the Xin'anjiang model calibrated in step 2.1 to obtain a long series of daily runoff simulations, and divide the simulations into three types of daily runoff series with different magnitudes. The method for dividing the simulation series into three types of daily runoff series of different magnitudes is as follows: calculate the 35% quantile and 65% quantile of the daily runoff series, and divide the daily runoff series into the first type below the 35% quantile, the second type between the 35% quantile and the 65% quantile, and the third type above the 65% quantile.

[0039] S2.4. Based on three types of daily runoff series of different magnitudes, long-short-term memory neural network models are constructed to correct the first simulated daily runoff process and reduce the hydrological model error caused by human activities.

[0040] This embodiment constructs a long short-term memory neural network (LSTM) model with a three-layer neural network architecture to generalize the regulating and storage effects of dams, reservoirs, or water diversion projects on watersheds and improve the accuracy of hydrological simulations. This embodiment uses a neural network interval simulation mean method to independently run the neural network model multiple times and take the average value as the final simulation result to reduce uncertainty.

[0041] To solve the problem of nonlinear autoregressive exogenous input model (NARX) dynamic neural network in deep learning process (number of hidden layers) To address the exploding and vanishing gradient problems caused by the NARX neural network (layer 2), the LSTM (Long Short-Term Memory) neural network enhances the long-term memory capacity of the NARX neural network by introducing storage units—namely, input gates, forget gates, internal feedback connections, and output gates—into the hidden layers of the NARX neural network. This allows the network to selectively remember current information or forget past information (such as the rainfall-runoff mapping). In short, the LSTM neural network replaces each hidden layer in the NARX dynamic neural network with a storage unit with memory capabilities, referred to as an LSTM unit. Its input and output layers are the same as those of the NARX dynamic neural network.

[0042] Based on the long short-term memory neural network (LSTM) model, the second simulated runoff series is obtained after correction, and the expression is: ](2); Where, express t The second simulated runoff after time correction, Indicates that the Xinanjiang model is t The first simulated daily runoff at time, Indicates that the Xinanjiang model is t- The first simulated daily runoff at time 1, N represents the lag time determined in step 2.2; F LSTM Represents an LSTM model.

[0043] Furthermore, the LSTM model is trained using the conventional technique in this field, the minimum batch gradient descent method.

[0044] Based on the above steps, a method for obtaining a second simulated runoff was constructed by using the Xin'anjiang model to simulate the first runoff, and further correcting it using a response LSTM model according to the type of runoff on each day. In this embodiment, the combination of the Xin'anjiang model, runoff type classification, and LSTM model is referred to as a deep learning model that considers hydrological processes.

[0045] Step 3: Input the long series of meteorological observation data obtained in step 1 into the deep learning model considering the hydrological process established in step 2 to reconstruct the long series of daily runoff process of the basin with scarce data. In this embodiment, the reconstructed long series of daily runoff is approximated to the measured runoff series.

[0046] A long series of daily runoff processes for a watershed with scarce data is reconstructed. The 80% quantile of the daily runoff is used as a threshold, and the daily runoff exceeding the threshold is identified and used as the flood event magnitude. For daily runoff that continuously exceeds the threshold, the average of the daily runoff during this period is taken as the flood event magnitude. However, when the calculated flood magnitude exceeds the 95% quantile of the daily runoff, the flood event is identified as an outlier, and the 95% quantile of the daily runoff is taken as the flood magnitude of the event. That is, for a particular flood event, the flood magnitude is calculated using the following formula: ; Where: is the flood magnitude of the event; and represent the 80% and 95% quantiles of the daily runoff series respectively; MM is the number of days in the flood event that exceeded the 80% quantile of the daily runoff, where ii =1,…,MM; iiThe flood event ii day.

[0047] Step 4: Based on meteorological data and extracted flood events, K artificial intelligence models are used to build a flood simulation model and calibrate the model parameters.

[0048] The following sub-steps are included: S4.1. Based on satellite remote sensing inversion data, in this embodiment, based on the land surface product of the fifth generation atmospheric reanalysis dataset (ERA5-Land) of the European Centre for Medium-Range Weather Forecasts, hourly meteorological and hydrological observation data are collected, specifically including 2m air temperature, air pressure, dew point temperature, snowfall, precipitation, wind speed, shortwave radiation, and longwave radiation. The daily average 2m air temperature, daily average dew point temperature, daily average snowfall, daily precipitation, daily average air pressure, daily wind speed, daily shortwave radiation, and daily longwave radiation are obtained through time scale conversion.

[0049] S4.2. Calculate specific humidity and relative humidity based on the Clausius-Clapeyron thermodynamic equation.

[0050] The Clausius-Clapeyron thermodynamic equation is defined as follows: (3); Where: is the first integration constant, which is 273.16 K; is the second integral constant, which is 611 Pa; is the latent heat of vaporization constant, take ; is the water vapor gas constant, take ; T is the air temperature, which is the input variable of the Clausius-Clapeyron thermodynamic equation; The 2-meter air temperature and dew point temperature of the meteorological and hydrological observation samples of each grid point in the study basin are substituted as input variables into the Clausius-Clapeyron thermodynamic equation to calculate the relative humidity of each grid point. RH , , where T dew is the dew point temperature, T 2m The air temperature at 2m; Specific humidity q is the ratio of water vapor mass to total air mass mass, using the ground pressure in the reanalysis data p And dew point temperature, the formula is as follows: (4); S4.3. Use the Thiessen polygon method to derive the average meteorological variables of the basin with scarce data. Combine the Honghu water events in the basin with the meteorological variables of the basin. The meteorological variables include the average daily temperature, specific humidity, relative humidity, snowfall, precipitation, wind speed, shortwave radiation, and longwave radiation on the day of the flood event and 1-3 days before. That is, for each flood event, 8×4=32 meteorological data are involved.

[0051] A random forest model was used to screen key meteorological factors. The random forest model can rank the importance of each input variable to the simulation variables. Therefore, based on this random forest model, important factors influencing floods were selected. In this example, a threshold of 25% was set. That is, after ranking the eight variables, the top eight variables were selected as key factors.

[0052] S4.4. Based on the reanalysis dataset, in this embodiment, based on the EAR5-Land reanalysis dataset, the key factors screened in step 4.3 are used as input, and the flood magnitude is used as the target variable to construct 15 artificial intelligence models, and the 15 calibrated artificial intelligence models are obtained through optimization training using the gradient descent method.

[0053] The 15 artificial intelligence models considered in this embodiment include: linear regression, logistic regression, decision tree, random forest, support vector machine, K-nearest neighbor algorithm, naive Bayes classifier, neural network, convolutional neural network, recurrent neural network, long short-term memory network, generative adversarial network, Transformer, XGBoost and Markov decision process model.

[0054] Step 5: Based on the simulation results of 15 flood simulation models and the measured flood magnitude, the Bayesian pattern averaging method is used to deduce the weight of each model, thereby constructing an artificial intelligence-Bayesian coupling model.

[0055] The following sub-steps are included: S5.1. Use the meteorological data of the reanalysis dataset in the satellite remote sensing inversion data as input, that is, use the meteorological data of ERA5-Land as input to drive the K artificial intelligence models suitable for flood magnitude simulation constructed in step 4 to obtain the simulated magnitude series of flood events in the basin with scarce data.

[0056] S5.2. Based on the simulation magnitude results of 15 artificial intelligence models and the measured flood index, the probability density function of the simulated flood magnitude was constructed according to the Bayesian total probability formula. Let S be the simulated flood magnitude, R = [D, O] represents the model input data, where D is the simulation series of each artificial intelligence model during the training period, and O is the measured series. The output results of 15 different artificial intelligence models are obtained by the Bayesian total probability formula. The probability density function of S is as follows: (5); Where: For the k An artificial intelligence model for flood magnitude simulation The probability density function of the simulated value S under the given data R; For a given training data R k The posterior probability density function of an artificial intelligence model suitable for flood magnitude simulation is given by: K represents the number of artificial intelligence models suitable for flood magnitude simulation. In this embodiment, K=15.

[0057] S5.3. Determine the corresponding weights based on the relative contributions of each AI model to simulate flood magnitude, thereby establishing a Bayesian pattern average correction model. The specific operations are as follows: First, the flood magnitude observation series and the simulation series obtained by each artificial intelligence model are transformed into normality using the Box-Cox function. Then, the estimation results of multiple models are weighted averaged based on the normal linear distribution assumption: (6); Where: Indicates the mean , the variance is Normal distribution; E represents the expected value of the function, For the k Weights for an artificial intelligence model suitable for flood magnitude simulation.

[0058] S5.4. Input the downscaled long series meteorological data set into the 15 artificial intelligence models suitable for flood magnitude simulation constructed in step 4 respectively, and then input the simulation results of K artificial intelligence models suitable for flood magnitude simulation into the Bayesian pattern averaging model, and obtain high-precision flood simulation results through the weighted averaging method of preferred weights.

[0059] In this embodiment, the coupling method of artificial intelligence and Bayesian model averaging model is called building an artificial intelligence-Bayesian coupling model, and the input of the coupling model is meteorological variables.

[0060] Step 6: Combine the Copula function to construct the distribution function of the measured flow and the forecast flow, and use the artificial intelligence-Bayesian coupling model to carry out flood forecasting.

[0061] The following steps are involved: S6.1. Obtain meteorological forecast information for data-scarce basins from the European Centre for Medium-Range Weather Forecasts (ECMWF) or the China Meteorological Administration (CMA), including daily average temperature, specific humidity, relative humidity, snowfall, precipitation, wind speed, shortwave radiation, and longwave radiation for the next 1-7 days.

[0062] S6.2. Use the Thiessen polygon method to calculate the regional average weather forecast information for the watershed with scarce data.

[0063] S6.3. The reanalysis meteorological data of the past three days were obtained from ERA5-Land, and the Thiessen polygon method was used to obtain the regional average reanalysis meteorological data of the data-scarce basin.

[0064] S6.4. Use the regional average weather forecast information for the next day of the scarce data basin obtained in step 6.2 and the regional average reanalysis weather data for the past 1 to 3 days of the scarce data basin obtained in step 6.3 as input to drive the artificial intelligence-Bayesian coupling model established in step 5 to obtain the flood forecast value for the next day.

[0065] S6.5. Using steps S6.1 to S6.4, use weather forecast information with a longer future forecast period as input to obtain flood forecast values for the next 2-3 days.

[0066] S6.6. Use the retrospective forecasting method to construct the joint distribution function of flood observations and flood forecasts over the historical period: make H k 、 S k Respectively represent k The actual flood magnitude at each moment and the deterministic flood forecast magnitude, h k 、 s k Then H k 、 S k The realized value of and They are H k 、 S k The marginal distribution function of and The corresponding H k 、 S k The probability density function of , according to Sklar's theorem, there exists a n -Copula function C , so that the following equation holds: (7); Where: Copula function parameters; is the density function of the two-dimensional Copula function; Represents the joint distribution function of the actual flood magnitude and the predicted magnitude.

[0067] is the Copula function parameter, often using the Kendall rank correlation coefficient Please solve.

[0068] Use Gumbel-Hougaard, Clayton and Frank Copula functions from the common Archimedean Copula function family to construct H k 、 S k The joint distribution of , and its related mathematical expressions are detailed in Table 1.

[0069]

[0070] S6.7. Derivation of the posterior distribution function of flood magnitude based on the Copula function and probabilistic forecasting; Given forecast value s k ,exist S k = s k hour, H k The conditional distribution function of for: (8); H k The density function is expressed as: (9); and When the deterministic forecast flood magnitude information is given, the measured flood magnitude h k The posterior probability distribution function and posterior probability density function are used to describe the uncertainty of the forecast; the expected value of the posterior probability density function is taken as the post-processing deterministic forecast result, and the confidence level is obtained at the same time The probability forecast interval of the flood magnitude under the condition of , and the expected value of the posterior probability density function of the measured flood magnitude h ke Solve the following equation: (10).

[0071] when When H k The confidence level is 90%. H kThe lower and upper confidence limits are taken as the 5% and 95% quantiles respectively to quantitatively describe the uncertainty of flood forecasting, and the information within the interval is used as the probabilistic forecast result.

[0072] Figure 4 This is a schematic diagram of the probabilistic flow forecast process and the 90% confidence forecast interval provided by the present invention. The ensemble forecast flow represented by the dotted line in the figure is the posterior expected value in this method.

[0073] Example 2: This embodiment provides a flood probability forecasting device 200 based on artificial intelligence, such as Figure 5 As shown, the method for executing the flood probability forecasting method described in Example 1 includes: A first determination module 202 is configured to construct a statistical downscaling model based on ground observation data and satellite remote sensing inversion data using an isorthogonal correction method, thereby obtaining a long series of precipitation and temperature observation data; An analysis module 204 is used to establish a watershed hydrological model based on a short series of runoff observation data and downscaled meteorological data in a watershed with scarce data; Building module 204, for using a random forest model to optimize key physical factors affecting runoff, and further using a long short-term memory model for correction, thereby building a deep learning model that considers hydrological processes; The second determination module 208 inputs the long series of meteorological observation data obtained by downscaling into a deep learning model that considers hydrological processes, reconstructs the long series of daily runoff processes in the watershed with scarce data, extracts flood events using an ultra-quantitative sampling method, and constructs an artificial intelligence-Bayesian coupled model; The prediction module 210 constructs the distribution function of the measured flow and the forecast flow in combination with the Copula function, and uses the artificial intelligence-Bayesian coupling model to carry out flood forecasting, and uses the Bayesian probability forecast processor based on the Copula function to quantify the forecast uncertainty and realize the flood probability forecast.

[0074] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A flood probability forecasting method based on artificial intelligence, characterized in that: The following steps are involved: Step 1: Based on ground observation data and satellite remote sensing inversion data, a statistical downscaling model is constructed using the equal-rate correction method. The downscaled long-series meteorological observation data, including long-series precipitation and temperature observation data, are obtained through the statistical downscaling model. Step 2: Based on the short-series runoff observation data and the downscaled long-series meteorological observation data of the data-scarce basin, a basin hydrological model is established. The long-short-term memory model is used for calibration to construct a deep learning model that takes into account the hydrological process. Step 3: Input the long series of meteorological observation data obtained in step 1 into the deep learning model considering the hydrological process established in step 2, reconstruct the long series of daily runoff processes in the basin with scarce data, and extract flood events using the super-quantitative sampling method; Step 4: Based on the meteorological data and the extracted flood events, K artificial intelligence models suitable for flood magnitude simulation are used to construct a flood simulation model, and the model parameters are calibrated; Step 5: Based on the simulation results of the K flood simulation models and the measured flood magnitude, the Bayesian pattern averaging method is used to deduce the weight of each model, thereby constructing an artificial intelligence-Bayesian coupling model; Step 6: Combine the Copula function to construct the distribution function of the measured flow and the predicted flow, and use the artificial intelligence-Bayesian coupling model to carry out flood forecasting.

2. The artificial intelligence-based flood probability forecasting method according to claim 1, characterized in that: The step 1 includes the following sub-steps: S1.

1. Collect ground observation data from ground stations in data-scarce areas. Ground observation data include limited meteorological observation data and hydrological observation series datasets. Satellite remote sensing inversion data include satellite-derived precipitation products and reanalysis datasets. S1.

2. Use the equal-rate correction method to construct a statistical downscaling model and correct the precipitation magnitude and temperature simulation bias on a monthly basis; The isorelative correction method assumes that the monthly deviations between satellite-retrieved precipitation and reanalysis temperature and ground-based observations are consistent over time. First, a correction factor is calculated for each month based on ground-based observations, and then the factor is applied to the long series of simulated data sets for the same month. The precipitation deviation ratio or absolute temperature deviation of the month is calculated based on the observation data and satellite remote sensing inversion data, and the corresponding correction factors are obtained to correct the satellite inverted precipitation series and reanalysis temperature series using the following formula: （1）； Where: N Indicates site m Total number of observation days per month; i Represents a daily precipitation or temperature series; Indicates that the satellite-retrieved precipitation m The original data series for the month, Indicates the corrected satellite-retrieved precipitation in the m Monthly series; Indicates that the satellite-retrieved precipitation m Month i The original data of the day, Indicates that the weather station is m Month i Daily observed precipitation data; Reanalysis temperature in the m The original data series for the month, The corrected reanalysis temperature product is m Monthly series; Reanalysis temperature product in the m Month i The original data of the day, Indicates that the weather station is m Month i Observed temperature data for the day.

3. The artificial intelligence-based flood probability forecasting method according to claim 1, characterized in that: The step 2 includes the following sub-steps: S2.

1. For basins with scarce data, the Xin'an River model is constructed using short-term runoff observation data and long-term meteorological observation data from the same period after statistical downscaling, and the model parameters are calibrated. S2.

2. Based on the calibrated Xin'anjiang model, the first simulated daily runoff process is simulated. By statistically analyzing the first simulated daily runoff process and the measured daily runoff process, the lag time affecting the measured daily runoff is determined. N ; S2.

3. Use the long series of precipitation and temperature observations corrected in step 1 to drive the Xin'anjiang model calibrated in step 2.1 to obtain a long series of daily runoff simulations, and divide the simulations into three types of daily runoff series with different magnitudes. S2.

4. Based on the three types of daily runoff series with different magnitudes, a long-short-term memory neural network model is constructed to correct the first simulated daily runoff process. After correction, the second simulated runoff series is obtained, which is expressed as: ]（2）； Where, express t The second simulated runoff after time correction, Indicates that the Xinanjiang model is t The first simulated daily runoff at time, Indicates that the Xinanjiang model is t- The first simulated daily runoff at time 1, N represents the lag time determined in step 2.2; F LSTM Represents an LSTM model.

4. The artificial intelligence-based flood probability forecasting method according to claim 3, characterized in that: In step S2.3, the method for dividing the simulation series into three types of daily runoff series of different magnitudes is: calculating the 35% quantile and 65% quantile of the daily runoff series, and dividing the daily runoff series into a first type below the 35% quantile, a second type between the 35% quantile and the 65% quantile, and a third type above the 65% quantile.

5. The artificial intelligence-based flood probability forecasting method according to claim 1, characterized in that: In step 3, a long series of daily runoff processes in the watershed with scarce data is reconstructed, and the 80% quantile of the daily runoff is used as a threshold to identify daily runoffs exceeding the threshold and use it as the flood event magnitude; for daily runoffs that continuously exceed the threshold, the average of the daily runoff during this period is taken as the flood event magnitude; when the calculated flood magnitude exceeds the 95% quantile of the daily runoff, the flood event is identified as an outlier, and the 95% quantile of the daily runoff is taken as the flood magnitude of the event.

6. The artificial intelligence-based flood probability forecasting method according to claim 1, characterized in that: The step 4 includes the following sub-steps: S4.

1. Based on satellite remote sensing inversion data, collect hourly meteorological and hydrological observation data, including 2-meter air temperature, air pressure, dew point temperature, snowfall, precipitation, wind speed, shortwave radiation, and longwave radiation. Obtain daily average 2-meter air temperature, daily average dew point temperature, daily snowfall, daily precipitation, daily average air pressure, daily wind speed, daily shortwave radiation, and daily longwave radiation through time scale conversion. S4.

2. Calculate specific humidity and relative humidity based on the Clausius-Clapeyron thermodynamic equation; The Clausius-Clapeyron thermodynamic equation is defined as follows: (3)； Where: is the first integration constant, which is 273.16 K; is the second integral constant, which is 611 Pa; is the latent heat of vaporization constant, take ; is the water vapor gas constant, take ; T is the air temperature, which is the input variable of the Clausius-Clapeyron thermodynamic equation; The 2-meter air temperature and dew point temperature of the meteorological and hydrological observation samples of each grid point in the study basin are substituted as input variables into the Clausius-Clapeyron thermodynamic equation to calculate the relative humidity of each grid point. RH , , where T dew is the dew point temperature, T 2m The air temperature at 2m; Specific humidity q is the ratio of water vapor mass to total air mass mass, using the ground pressure in the reanalysis data p And dew point temperature, the formula is as follows: (4)； S4.

3. Use the Thiessen polygon method to derive average meteorological variables for basins with scarce data. Combine the Honghu flood events in the basin with the basin's meteorological variables. The meteorological variables include the average daily temperature, specific humidity, relative humidity, snowfall, precipitation, wind speed, shortwave radiation, and longwave radiation on the day of the flood event and 1-3 days before. Use a random forest model to screen key meteorological factors. S4.

4. Based on the reanalysis dataset, using the key meteorological factors selected in step S4.3 as input and flood magnitude as the target variable, construct K artificial intelligence models suitable for flood magnitude simulation.

7. The artificial intelligence-based flood probability forecasting method according to claim 6, characterized in that: In step S4.4, the K artificial intelligence models suitable for flood magnitude simulation include: linear regression, logistic regression, decision tree, random forest, support vector machine, K nearest neighbor algorithm, naive Bayes classifier, neural network, convolutional neural network, recurrent neural network, long short-term memory network, generative adversarial network, Transformer, XGBoost and Markov decision process model.

8. The artificial intelligence-based flood probability forecasting method according to claim 1, characterized in that: The step 5 includes the following sub-steps: S5.

1. Use meteorological data from the reanalysis dataset of satellite remote sensing inversion data as input to drive the K artificial intelligence models suitable for flood magnitude simulation constructed in step 4 to obtain a simulated magnitude series of flood events in the data-scarce basin. S5.

2. Based on the simulated flood magnitude results and measured flood indices of K AI models suitable for flood magnitude simulation, a probability density function of the simulated flood magnitude is constructed according to the Bayesian total probability formula. Let S be the simulated flood magnitude, and R = [D, O] represent the model input data, where D is the simulation series of each AI model during the training period, and O is the measured series. The output results of K different artificial intelligence models are obtained by the Bayesian total probability formula. The probability density function of S is as follows: （5）； Where: For the k An artificial intelligence model for flood magnitude simulation The probability density function of the simulated value S under the given data R; For a given training data R k The posterior probability density function of the artificial intelligence models suitable for flood magnitude simulation, K represents the number of artificial intelligence models; S5.

3. Determine the corresponding weights based on the relative contributions of each AI model to simulate flood magnitude, thereby establishing a Bayesian pattern average correction model. The specific operations are as follows: First, the flood magnitude observation series and the simulation series obtained by each artificial intelligence model are transformed into normality using the Box-Cox function. Then, the estimation results of multiple models are weighted averaged based on the normal linear distribution assumption: （6）； Where: Indicates the mean , the variance is Normal distribution; E represents the expected value of the function, For the k The weights of an AI model for flood magnitude simulation; S5.

4. Input the downscaled long series meteorological data set into the K artificial intelligence models suitable for flood magnitude simulation constructed in step 4 respectively, and then input the simulation results of the K artificial intelligence models suitable for flood magnitude simulation into the Bayesian pattern averaging model, and obtain high-precision flood simulation results through the weighted averaging method of preferred weights.

9. The artificial intelligence-based flood probability forecasting method according to claim 1, characterized in that: The step 6 includes the following steps: S6.

1. Obtain weather forecast information for watersheds where data is scarce, including daily average temperature, specific humidity, relative humidity, snowfall, precipitation, wind speed, shortwave radiation, and longwave radiation for the next 1-7 days; S6.

2. Use the Thiessen polygon method to calculate the regional average weather forecast information for the watershed where data are scarce; S6.

3. Obtain the reanalysis meteorological data for the past three days from the reanalysis dataset and use the Thiessen polygon method to obtain the regional average reanalysis meteorological data for the data-scarce basin. S6.

4. Use the regional average weather forecast for the next day for the data-scarce basin obtained in step 6.2 and the regional average reanalysis weather data for the past 1 to 3 days obtained in step 6.3 as input to drive the artificial intelligence-Bayesian coupling model established in step 5 to obtain the flood forecast for the next day. S6.5, using steps S6.1 to S6.4, using weather forecast information with a longer future horizon as input, thereby obtaining flood forecast values for the next 2-3 days; S6.

6. Use the retrospective forecasting method to construct the joint distribution function of flood observations and flood forecasts over the historical period: make H k 、 S k Respectively represent k The actual flood magnitude at each moment and the deterministic flood forecast magnitude, h k 、 s k Then H k 、 S k The realized value of and They are H k 、 S k The marginal distribution function of and The corresponding H k 、 S k The probability density function of , according to Sklar's theorem, there exists a n -Copula function C , so that the following formula holds: (7)； Where: Copula function parameters; is the density function of the two-dimensional Copula function; The joint distribution function representing the actual flood magnitude and the predicted magnitude; S6.

7. Derivation of the posterior distribution function of flood magnitude based on the Copula function and probabilistic forecasting; Given forecast value s k ,exist S k = s k hour, H k The conditional distribution function of for: (8)； H k The density function is expressed as: (9)； and When the deterministic forecast flood magnitude information is given, the measured flood magnitude h k The posterior probability distribution function and posterior probability density function are used to describe the uncertainty of the forecast; the expected value of the posterior probability density function is taken as the post-processing deterministic forecast result, and the confidence level is obtained at the same time The probability forecast interval of the flood magnitude under the condition of , and the expected value of the posterior probability density function of the measured flood magnitude h ke Solve the following equation: (10)。 10. A flood probability forecasting device based on artificial intelligence, characterized in that: The method for executing the flood probability forecasting method according to any one of claims 1 to 9 comprises: The first determination module is used to construct a statistical downscaling model based on ground observation data and satellite remote sensing inversion data using the equal-rate correction method to obtain long-series precipitation and temperature observation data; An analysis module is used to build a basin hydrological model based on short-series runoff observation data and downscaled meteorological data in data-scarce basins; A module was built to use a random forest model to select key physical factors affecting runoff, and further use a long-short-term memory model for correction, thereby constructing a deep learning model that considers hydrological processes. The second determination module inputs the long series of meteorological observation data obtained by downscaling into a deep learning model that considers hydrological processes, reconstructs the long series of daily runoff processes in the basin with scarce data, extracts flood events using the ultra-quantitative sampling method, and constructs an artificial intelligence-Bayesian coupled model; The prediction module combines the Copula function to construct the distribution function of measured flow and forecast flow, and uses the artificial intelligence-Bayesian coupling model to carry out flood forecasting.

Images