High-cold basin rain-snow-ice mixed flood forecasting method based on physical guidance deep learning

Abstract

The physical guidance deep learning based high-cold basin rain-snow-ice mixed flood forecasting method belongs to the technical field of cross of hydrology and meteorology forecasting, water conservancy engineering and artificial intelligence. The method comprises the following steps: calculating a multi-dimensional feature set, extracting effective physical coefficients based on physical concepts, nonlinear mapping of a long short-term memory neural network model actual runoff upper limit and a flood runoff coefficient, and deep fusion of a double-channel asymmetric Nash parallel confluence module, and finally constructing a physical guidance deep learning based high-cold basin rain-snow-ice mixed flood forecasting model to realize the whole process prediction of the high-cold basin rain-snow-ice mixed flood. The method can realize the whole process transparent prediction from meteorological forcing to runoff component decoupling and then asymmetric confluence evolution, significantly improve the physical consistency and generalization robustness of complex mixed flood prediction under extreme flood, significantly improve the reliability and explainability of complex mixed flood prediction.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of hydrological and meteorological forecasting, water conservancy engineering and artificial intelligence, and relates to a method for forecasting mixed rain, snow and ice floods in cold water basins based on physical-guided deep learning. Background Technology

[0002] Floods are among the most destructive natural disasters globally, posing a significant threat to human life, property, and socio-economic infrastructure. In recent years, against the backdrop of global warming, the accelerated melting of glaciers and snow in high-altitude mountainous areas has led to significant changes in hydrological conditions. Complex floods, driven by a mixture of multiple forcing signals including rainfall, snowmelt, and ice melt, have seen a dramatic increase in frequency and uncertainty. This is particularly true for inland arid cities (such as Urumqi, China), where mountain runoff is not only a natural disaster requiring immediate defense but also the lifeline for maintaining the local oasis economy and ecosystem. In this context of extreme water scarcity, floods possess the dual attributes of both "disaster" and "resource." The inherent contradiction between flood control and water storage places extremely stringent demands on the timeliness and accuracy of reservoir flood control scheduling: even a slight error in forecasting could lead to dam failure or the loss of precious flood resources. Therefore, constructing high-precision, high-timeliness hourly flood forecasting models is not only a highly challenging scientific problem but also an urgent engineering necessity.

[0003] However, achieving accurate forecasting of mixed rain and snow floods in high-altitude and cold mountainous areas faces complex scientific challenges. While traditional distributed physical-hydrological models possess clear causal physical mechanisms, their construction often relies on extensive hydrological and meteorological observations and high-resolution underlying surface data. High-altitude and cold mountainous areas are typically data-scarce regions with sparsely distributed stations and short observation sequences. Furthermore, physical models involve numerous parameters requiring calibration, easily leading to severe "heterogeneous but consistent spectrum" problems in data-scarce areas. In addition, static artificial parameters often struggle to adapt to rapid changes in non-stationary processes such as snowmelt under extreme weather conditions. In contrast, data-driven deep learning models, such as Long Short-Term Memory (LSTM) networks, do not require explicit construction of complex partial differential equations. With their powerful nonlinear feature extraction and long-range temporal memory capabilities, they have demonstrated superior fitting accuracy in hydrological simulations compared to traditional physical models in recent years, becoming a highly favored tool in flood forecasting.

[0004] While deep learning models excel at fitting observed data, they suffer from fatal limitations in practical flood control applications: a lack of physical consistency and interpretability. Purely data-driven models, acting as "data-hungry" black boxes, are highly susceptible to spurious correlations and overfitting in small-sample, high-altitude mountainous regions. More seriously, because deep learning models focus solely on the statistical mapping between input and output, ignoring the inherent conservation laws of the hydrological cycle, they are prone to producing absurd results that violate mass conservation or temporal logic when faced with extreme weather events outside of historical distributions. This unreliability is completely unacceptable in the highly sensitive flood control scheduling of reservoirs.

[0005] To address the above issues, extensive research has been conducted both domestically and internationally, resulting in several solutions. For example, Chinese invention patent (application number 2025103375515) provides an interpretable AI-based flood forecasting method based on watershed runoff generation and confluence mechanisms. This method utilizes deep learning integrated with physics to forecast floods, avoiding the difficulties of parameter calibration in traditional models and effectively improving the interpretability of deep learning in flood forecasting from the perspective of runoff generation and confluence mechanisms. Another Chinese invention patent (application number 2022111928475) provides a rainfall-driven deep learning flood forecasting method based on runoff generation and confluence processes. This method also employs deep learning integrated with physics, forecasting floods from a rainfall-driven perspective, while simultaneously enhancing the interpretability of the deep learning flood forecasting method by incorporating runoff generation and confluence mechanisms.

[0006] While existing technical solutions employ deep learning methods to avoid the difficulties in calibrating traditional models, there is still a gap in research on deep learning forecasting methods for complex mixed floods involving rain, snow, and ice. Furthermore, the process of combining flood generation and confluence mechanisms with flood forecasting, which suffers from low interpretability, remains unclear. Summary of the Invention

[0007] To address the shortcomings of existing technologies, such as the difficulty in determining parameters of traditional physical models and the lack of physical mechanisms and interpretability in deep learning models, this invention provides a method for forecasting mixed rain, snow, and ice floods in high-altitude watersheds based on physical-guided deep learning. This invention embeds hydrological energy balance, mass conservation, and dynamic evolution mechanisms into a neural network architecture, achieving transparent decoupling and asymmetric confluence estimation of runoff from rain, snow, and ice sources.

[0008] To achieve the above objectives, the specific technical solution adopted by the present invention is as follows: A method for forecasting mixed rain, snow, and ice floods in high-altitude cold-climate watersheds based on physical-guided deep learning, comprising the following steps: Step S1: Obtain multi-source data of the target watershed, preprocess the multi-source data, and form a multi-dimensional feature set based on this data. The multi-dimensional feature set includes a static pre-processing feature set and a dynamic time-series feature set; specifically: Step S1.1: Obtain multi-source data for the target watershed. The multi-source data includes rainfall data, temperature data, and runoff data. Specifically: the rainfall data originates from measured meteorological station data and multi-source fused precipitation datasets, including but not limited to daily rainfall data, flood season rainfall data, and hourly rainfall data, with a temporal resolution of at least daily; the temperature data originates from measured meteorological station data and multi-source fused temperature data, including but not limited to daily average temperature data, daily maximum and minimum temperature data; the runoff data originates from measured hydrological station data and multi-source fused runoff data, including but not limited to daily runoff data and hourly runoff data during flood events.

[0009] Step S1.2 involves preprocessing the multi-source data obtained in Step S1.1. Linear interpolation is used to uniformly resample the rainfall and runoff data at different time scales to obtain continuous daily rainfall data P, daily runoff data Q, and hourly rainfall data P during the flood period. t Hourly runoff data Q t The temperature data was resampled using an asymmetric cosine interpolation method to obtain smooth hourly temperature data T.

[0010] Step S1.3: Based on the data preprocessed in step S1.2, a multidimensional feature set is constructed, which includes a dynamic time-series feature set and a static early-stage feature set.

[0011] The static pre-flood feature set is used to characterize the basin's baseline state before a single flood event, specifically including the pre-flood snow storage value Sd, pre-flood impact rainfall Pa, pre-flood melt impact Sia, and the previous day's temperature increment ΔT. The pre-flood snow storage value is calculated daily from the beginning of the year to the day before the flood using a degree-day factor method with a set temperature threshold. The pre-flood impact rainfall is also calculated daily from the beginning of the year to the day before the flood. The pre-flood melt impact is calculated using the API model, with attenuation factors set according to basin characteristics. The previous day's temperature increment is the temperature difference between the day before the flood and the day the flood begins. The temperature threshold range should be realistic, for example, 0℃ to 2℃. When rainfall occurs, precipitation below 0℃ is pure solid snowfall, and above 2℃ is pure liquid precipitation. Within the range, linear interpolation is used to calculate the ratio of solid snowfall to liquid precipitation.

[0012] The dynamic time series feature set is used to characterize the meteorological driving forces during the flood process, specifically including time-series rainfall, time-series temperature, time-series runoff, and time-series snow storage value divided into each calculation time step (such as hourly).

[0013] Step S2: Based on the multi-dimensional feature set from Step S1, calculate the effective physical coefficients for rain, snow, and ice respectively. Construct a long short-term memory neural network model and introduce the effective physical coefficients to map the actual runoff upper limit. Simultaneously, predict the runoff coefficient of each flood event using the multi-dimensional feature set. Multiply the actual runoff upper limit by the runoff coefficient of each flood event to obtain the runoff for each time period. Specifically: Step S2.1: Calculate the effective physical coefficients. Considering the complex mixed flood characteristics of rain, snow, and ice in high-altitude cold basins, based on the time-period rainfall and temperature obtained in step S1, calculate the effective physical coefficients, including the effective rainfall coefficient P. eff Effective snow coefficient S eff and effective ice coefficient I eff These three effective physical coefficients are calculated based on hydrophysical concepts and are used to characterize the maximum runoff potential of rainwater, snow, and ice in a flood event. The specific calculation method is as follows: The effective rainfall coefficient is the representative coefficient that has the greatest impact on rainfall runoff. It is calculated by deducting the amount of solid snowfall, the actual losses due to plant interception and surface depression interception, and considering the infiltration effect under extreme rainstorm conditions, after defining the temperature threshold range of precipitation phase change. The temperature threshold range of precipitation phase change should be consistent with reality, such as a lower limit of 0℃ and an upper limit of 2℃.

[0014] The effective snow coefficient is the representative coefficient that most significantly affects snowmelt runoff, determined by the current snow storage value and temperature over a given period, and calculated using the degree-day factor method based on temperature. The effective ice coefficient is calculated using the day factor method and is a representative coefficient of the maximum melt flow generated by exposed glaciers under purely thermodynamic driving.

[0015] Step S2.2: Construct a long short-term memory network model.

[0016] A Long Short-Term Memory (LSTM) neural network model is constructed as the underlying computational architecture. The LSTM neural network model is a special type of recurrent neural network that internally contains memory units c. t And three control gating mechanisms, the three control gating mechanisms being input gate i t Forgotten Gate t Output gate o t Among them: memory unit c t Used to store long-term hydrological information in the sequence; input gate i t Controls the proportion of new meteorological information written to memory units at the current moment; forget gate f t Determine how much soil or snow history information from earlier periods to retain; Output gate o tThe influence of the memory unit state on the final hydrological output is controlled. This structure effectively solves the gradient vanishing problem that easily occurs in traditional networks, enabling the model to retain key hydrological information over long periods of time.

[0017] Step S2.3 introduces effective physical coefficients to map the actual runoff upper limit. The dynamic time series feature set obtained in step S1.3 and the effective physical coefficients obtained in step S2.1 are input into the LSTM neural network model constructed in step S2.2 for nonlinear mapping. Internally, the activation function calculates the actual runoff upper limit corresponding to each time step, which represents the physical limits of the three water sources, specifically including: rain runoff upper limit, snow runoff upper limit, and ice runoff upper limit.

[0018] Step S2.4: Predict the runoff coefficient of the flood event using a multi-dimensional feature set. The static pre-flood feature set obtained in step S1.3 is input into the LSTM neural network model to evaluate the initial surface dryness and wetness and water storage status of the entire flood event, thereby predicting the global runoff reduction ratio of the entire flood event, i.e., the runoff coefficient of the flood event, specifically including: rain runoff coefficient α, snow runoff coefficient β, and ice runoff coefficient γ.

[0019] Step S2.5: Calculate the runoff for each time period. Multiply the actual runoff upper limit of each time period obtained from step S2.3 by the corresponding flood runoff coefficient predicted in step S2.4 to obtain the effective net rainfall actually generated by each water source at each moment, i.e., the runoff for each time period, specifically including: time period rain runoff, time period snow runoff, and time period ice runoff.

[0020] Step S3: Based on the runoff generation for each time period obtained in Step S2, two Nash runoff models with different physical response characteristics are constructed for runoff calculation, including a fast Nash runoff model and a slow Nash runoff model. The runoff results of the two models are then linearly superimposed to obtain the total predicted runoff at the watershed outlet. Specifically: Step S3.1: Construct a fast Nash confluence model. Considering the physical characteristics of rainfall directly forming surface runoff, rapid confluence velocity, and sharp flood peaks, a fast Nash confluence model is constructed, which is a discrete causal convolutional module with a small time constant. The inter-period rainfall runoff calculated in step S2.5 is input into the fast Nash confluence model to calculate the simulated flow rate of the rainfall component at the watershed outlet, i.e., the simulated flow rate Q of the rainfall component. p : (1) in, This represents the number of linear reservoirs connected in series in the fast Nash confluence model; For the water storage calculation constant of a single reservoir in the fast Nash confluence model; It is a gamma function; For which moment?

[0021] Step S3.2: Construct a slow Nash confluence model. Considering the physical characteristics of snowmelt, which primarily flows through underground pores or seeps from the bottom of glaciers, exhibiting slow flow velocity, strong hysteresis effects, and gentle waveforms, a slow Nash confluence model is constructed, which is a discrete causal convolutional module with a large time constant. The time-period snow runoff and time-period ice runoff calculated in step S2.5 are simultaneously input into the slow Nash confluence model to calculate the simulated flow rate of the ablation component at the watershed outlet, thus obtaining the simulated flow rate Q of the ablation component. m : (2) in, This represents the number of linear reservoirs connected in series in the slow Nash confluence model; K m represents the water storage calculation constant for a single reservoir in the slow Nash confluence model.

[0022] Step S3.3, calculate the total predicted runoff. The simulated flow rate Q of the rainfall component obtained in step S3.1 is then used. p The simulated flow rate Q of the ablation component obtained in step S3.2 m Linear summation is performed at the same time points, and in actual forecasts, the flood rise flow is added to obtain the total forecast runoff Q at the basin outlet at each time point. sim .

[0023] Step S4: Based on the total predicted runoff output in Step S3 and the measured runoff data from hydrological stations obtained in Step S1, a multi-objective composite loss function is constructed. Iterative training is then performed using historical flood data to obtain a physical-guided deep learning flood forecasting model. Specifically: Step S4.1: Construct a multi-objective composite loss function. Extract the total predicted runoff Q obtained in step S3.3. sim Compared with the hourly runoff data Q obtained during the flood period from step S1.2 preprocessing t The Kling-Gupta efficiency coefficient (KGE) of both parameters is calculated as the main loss term to evaluate the goodness of fit of the overall process line. At the same time, the extreme value error of both parameters at the flood peak is calculated as the penalty auxiliary term. In addition, hydrophysical boundaries are set for the parameters output by the LSTM neural network model, including three runoff generation coefficients, fast and slow confluence parameters, and a boundary violation penalty term is constructed. The main loss term, penalty auxiliary term, and boundary violation penalty term are weighted to form a multi-objective composite loss function.

[0024] Step S4.2, iterative optimization of the physical-guided deep learning flood forecasting model. The error gradient is calculated using a multi-objective composite loss function, and all learnable parameters in the LSTM neural network model, the fast Nash confluence model, and the slow Nash confluence model are updated using the backpropagation algorithm. Steps 2.4 to 4.1 are repeated until the multi-objective composite loss function reaches its optimum, resulting in a fully calibrated physical-guided deep learning flood forecasting model, abbreviated as PG-LSTM flood forecasting model. Furthermore, the condition for the multi-objective composite loss function to reach its optimum is that the multi-objective composite loss function value is minimized.

[0025] Step S5: Based on the physically guided deep learning flood forecasting model trained in Step S4, input the multi-dimensional feature set of the flood event to be forecasted, and output the full-process flood forecast results and sub-item forecast results. Specifically: Step S5.1: Perform full-process forecast simulation. Obtain meteorological forecast data for the event to be forecasted and calculate a multi-dimensional feature set as in step S1. Input this feature set into the physical-guided deep learning flood forecasting model obtained in step S4. After forward propagation operations by the runoff generation module (step S2) and the confluence module (step S3), directly output the total forecast runoff process line for the entire future time period of the event to be forecasted.

[0026] Step S5.2: Output the sub-project forecast results. Based on the decoupling structure within the physical-guided deep learning flood forecast model, while outputting the total forecast runoff, the runoff for each time period obtained in step S2 and the simulated flow rate Q of the rainfall component output in step S3.1 are extracted separately. p The simulated flow rate Q of the ablation component obtained in step S3.2 m It calculates the proportion of each hydrophysical component in the total runoff and total runoff at each moment, realizing full-process transparency and interpretable forecasting of mixed rain, snow and ice floods in high-altitude cold basins.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) In step S1, the present invention uses linear interpolation and asymmetric cosine interpolation to preprocess the multi-source data of the target watershed and construct a multi-dimensional feature set. It integrates the static early feature set and the dynamic time series feature set, and comprehensively describes the hydrological state of the high-altitude watershed from two dimensions: long-term underlying surface drought evolution and short-duration meteorological forcing. It effectively eliminates the discontinuity and noise mutation of cross-source spatiotemporal scales, and lays a reliable data foundation for the calculation of runoff from multiple water sources such as rain, snow and ice.

[0028] (2) The present invention calculates the effective physical coefficients characterizing the maximum runoff potential of water sources through step S2, and uses a long short-term memory neural network model combined with multidimensional feature set nonlinear mapping to obtain the upper limit of actual runoff and the runoff coefficient of each flood. The two are multiplied one by one to obtain the runoff of each time period. The hydrological physical boundary depth is embedded in the deep learning calculation graph and the back propagation of the gradient is rigidly constrained, which overcomes the limitation of the traditional pure data-driven model that violates the causal logic of the hydrological system due to the "black box" mathematical fitting.

[0029] (3) In step S3, this invention classifies and constructs fast Nash confluence models and slow Nash confluence models in parallel to address the different response characteristics of direct surface runoff from rainfall and underground seepage from snowmelt. This describes the length of different water source flow paths and the differences in hysteresis, restores the complex interwoven waveform of mixed rain, snow and ice floods, and eliminates the velocity distortion and peak misalignment caused by traditional single confluence structures.

[0030] In summary, this invention constructs a physical-guided deep learning-based forecasting model for mixed rain, snow, and ice floods in cold water basins by calculating a multi-dimensional feature set, extracting effective physical coefficients based on physical concepts, nonlinearly mapping the actual runoff upper limit of the long short-term memory neural network model with the runoff coefficient of each flood event, and deeply integrating a dual-channel asymmetric Nash parallel confluence module. This model enables full-process forecasting of mixed rain, snow, and ice floods in cold water basins, significantly improving the reliability and interpretability of complex mixed flood forecasting. Attached Figure Description

[0031] Figure 1 This is an overall flowchart of the hybrid flood forecasting method based on physical guidance deep learning in an embodiment of the present invention; Figure 2 This is a comparison chart of the flood forecast flow process mainly based on heavy rainfall predicted by the PG-LSTM model at the Hero Bridge station in this embodiment of the invention. Figure 3 This is a comparison chart of the flood forecast flow process mainly based on snowmelt and ice melt in the PG-LSTM model of the Hero Bridge Station in this embodiment of the invention; Figure 4 This is a comparison chart of the flood forecast flow process mainly based on the mixture of heavy rain and snowmelt / ice melt in the PG-LSTM model of the Hero Bridge Station in this embodiment of the invention. Figure 5 This is a diagram showing the three runoff processes of rainfall, snowmelt, and ice melt for the PG-LSTM model forecast of floods mainly caused by heavy rain in an embodiment of the present invention at the Hero Bridge Station. Figure 6 This is a diagram showing the three runoff processes of rainfall, snowmelt, and icemelt in the PG-LSTM model forecast of floods mainly caused by snowmelt and icemelt at the Hero Bridge Station in this embodiment of the invention. Figure 7This is a diagram showing the three runoff processes of rainfall, snowmelt, and icemelt in the flood forecast by the PG-LSTM model at the Hero Bridge Station in this embodiment of the invention. Figure 8 This is a diagram showing the two confluence processes of rainfall and snowmelt / icemelt in the flood forecast by the PG-LSTM model at the Hero Bridge station in this embodiment of the invention. Figure 9 This is a diagram showing the two confluence processes of rainfall and snowmelt / icemelt in the PG-LSTM model forecast of floods mainly caused by snowmelt and icemelt at the Hero Bridge station in this embodiment of the invention. Figure 10 This is a diagram showing the two confluence processes of rainfall and snowmelt / icemelt in the flood forecast by the PG-LSTM model at the Hero Bridge station in this embodiment of the invention. Detailed Implementation

[0032] The specific embodiments of the present invention will be further described in detail below with reference to specific hydrological application scenarios. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0033] A method for forecasting mixed rain, snow, and ice floods in high-altitude, cold-climate watersheds based on physical-guided deep learning. The overall flowchart is as follows: Figure 1 As shown, it includes the following steps: Step S1: Taking the upper reaches of a river basin (a city in an arid inland area) as the target basin, acquire multi-source data for the target basin, preprocess the multi-source data, and form a multi-dimensional feature set based on this data. The multi-dimensional feature set includes a static pre-processing feature set and a dynamic time-series feature set; specifically: Step S1.1: Obtain multi-source data for the target watershed. The multi-source data includes rainfall data, temperature data, and runoff data. Specifically: the rainfall data originates from measured meteorological station data and multi-source fused precipitation datasets, including but not limited to daily rainfall data, flood season rainfall data, and hourly rainfall data, with a temporal resolution of at least daily; the temperature data originates from measured station data and multi-source fused temperature data, including but not limited to daily average temperature data, daily maximum and minimum temperature data; the runoff data originates from measured station data and multi-source fused runoff data, including but not limited to daily runoff data and hourly runoff data during flood events. The measured stations include the Hero Bridge Hydrological Station, the General Control Hydrological Station, the Houxia Camp Hydrological Station, and the No. 1 Glacier Hydrological Station.

[0034] Step S1.2 involves preprocessing the multi-source data obtained in Step S1.1. Linear interpolation is used to uniformly resample the rainfall and runoff data at different time scales to obtain continuous daily rainfall data P, daily runoff data Q, and hourly rainfall data P during the flood period. t Hourly runoff data Q tThe temperature data was resampled using an asymmetric cosine interpolation method to obtain smooth hourly temperature data T.

[0035] Step S1.3: Based on the data preprocessed in step S1.2, a multidimensional feature set is constructed, which includes a dynamic time-series feature set and a static early-stage feature set.

[0036] The static pre-flood feature set is used to characterize the basin's baseline state before a single flood event, specifically including the pre-flood snow storage value Sd, pre-flood impact rainfall Pa, pre-flood melt impact SIa, and the previous day's temperature increment ΔT. The pre-flood snow storage value is calculated daily from April 1st of the year to the day before the flood using a degree-day factor method with a set temperature threshold. The pre-flood impact rainfall is also calculated daily from April 1st of the year to the day before the flood. The pre-flood melt impact is calculated using the API model, with a decay factor set according to the basin's characteristics. The previous day's temperature increment is the temperature difference between the day before the flood and the day the flood begins. In this embodiment, the temperature threshold is 1℃.

[0037] The dynamic temporal feature set is used to characterize the meteorological driving forces during the flood process, specifically including hourly time-based rainfall, time-based temperature, time-based runoff, and time-based snow storage value.

[0038] Step S2: Based on the multi-dimensional feature set from Step S1, calculate the effective physical coefficients for rain, snow, and ice respectively. Construct a long short-term memory neural network model and introduce the effective physical coefficients to map the actual runoff upper limit. Simultaneously, predict the runoff coefficient of each flood event using the multi-dimensional feature set. Multiply the actual runoff upper limit by the runoff coefficient of each flood event to obtain the runoff for each time period. Specifically: Step S2.1: Calculate the effective physical coefficients. Considering the complex mixed flood characteristics of rain, snow, and ice in high-altitude cold basins, based on the time-period rainfall and temperature obtained in step S1, calculate the effective physical coefficients, including the effective rainfall coefficient P. eff Effective snow coefficient S eff and effective ice coefficient I eff These three effective physical coefficients are calculated based on hydrophysical concepts and are used to characterize the maximum runoff potential of rainwater, snow, and ice in a flood event. The specific calculation method is as follows: The effective rainfall coefficient is the representative coefficient that has the greatest impact on rainfall runoff. It is calculated by deducting solid snowfall from the temperature threshold range of precipitation phase transition, deducting the actual losses due to plant interception and surface depression interception, and considering the infiltration effect under extreme rainstorm conditions. The lower limit of the temperature threshold range of precipitation phase transition is 0℃ and the upper limit is 2℃.

[0039] The effective snow coefficient is the representative coefficient that most significantly affects snowmelt runoff, determined by the current snow storage value and temperature over a given period, and calculated using the degree-day factor method based on temperature. The effective ice coefficient is calculated using the day factor method and is a representative coefficient of the maximum melt flow generated by exposed glaciers under purely thermodynamic driving.

[0040] Step S2.2: Construct a long short-term memory network model.

[0041] A Long Short-Term Memory (LSTM) neural network model is constructed as the underlying computational architecture. The LSTM neural network model is a special type of recurrent neural network that internally contains memory units c. t And three control gating mechanisms, the three control gating mechanisms being input gate i t Forgotten Gate t Output gate o t Among them: memory unit c t Used to store long-term hydrological information in the sequence; input gate i t Controls the proportion of new meteorological information written to memory units at the current moment; forget gate f t Determine how much soil or snow history information from earlier periods to retain; Output gate o t The influence of the memory unit state on the final hydrological output is controlled. This structure effectively solves the gradient vanishing problem that easily occurs in traditional networks, enabling the model to retain key hydrological information over long periods of time.

[0042] Step S2.3 introduces effective physical coefficients to map the actual runoff upper limit. The dynamic time series feature set obtained in step S1.3 and the effective physical coefficients obtained in step S2.1 are input into the LSTM neural network model constructed in step S2.2 for nonlinear mapping. Internally, the activation function calculates the actual runoff upper limit corresponding to each time step, which represents the physical limits of the three water sources, specifically including: rain runoff upper limit, snow runoff upper limit, and ice runoff upper limit.

[0043] Step S2.4: Predict the runoff coefficient of the flood event using a multi-dimensional feature set. The static pre-flood feature set obtained in step S1.3 is input into the LSTM neural network model to evaluate the initial surface dryness and wetness and water storage status of the entire flood event, thereby predicting the global runoff reduction ratio of the entire flood event, i.e., the runoff coefficient of the flood event, specifically including: rain runoff coefficient α, snow runoff coefficient β, and ice runoff coefficient γ.

[0044] Step S2.5: Calculate the runoff for each time period. Multiply the actual runoff upper limit of each time period obtained from step S2.3 by the corresponding flood runoff coefficient predicted in step S2.4 to obtain the effective net rainfall actually generated by each water source at each moment, i.e., the runoff for each time period, specifically including: time period rain runoff, time period snow runoff, and time period ice runoff.

[0045] Step S3: Based on the runoff generation for each time period obtained in Step S2, two Nash runoff models with different physical response characteristics are constructed for runoff calculation, including a fast Nash runoff model and a slow Nash runoff model. The runoff results of the two models are then linearly superimposed to obtain the total predicted runoff at the watershed outlet. Specifically: Step S3.1: Construct a fast Nash confluence model. Considering the physical characteristics of rainfall directly forming surface runoff, rapid confluence velocity, and sharp flood peaks, a fast Nash confluence model is constructed, which is a discrete causal convolutional module with a small time constant. The inter-period rainfall runoff calculated in step S2.5 is input into the fast Nash confluence model to calculate the simulated flow rate of the rainfall component at the watershed outlet, i.e., the simulated flow rate Q of the rainfall component. p : (1) in, This represents the number of linear reservoirs connected in series in the fast Nash confluence model; For the water storage calculation constant of a single reservoir in the fast Nash confluence model; It is a gamma function; For which moment?

[0046] Step S3.2: Construct a slow Nash confluence model. Considering the physical characteristics of snowmelt, which primarily flows through underground pores or seeps from the bottom of glaciers, exhibiting slow flow velocity, strong hysteresis effects, and gentle waveforms, a slow Nash confluence model is constructed, which is a discrete causal convolutional module with a large time constant. The time-period snow runoff and time-period ice runoff calculated in step S2.5 are simultaneously input into the slow Nash confluence model to calculate the simulated flow rate of the ablation component at the watershed outlet, thus obtaining the simulated flow rate Q of the ablation component. m : (2) in, This represents the number of linear reservoirs connected in series in the slow Nash confluence model; K m represents the water storage calculation constant for a single reservoir in the slow Nash confluence model.

[0047] Step S3.3, calculate the total predicted runoff. The simulated flow rate Q of the rainfall component obtained in step S3.1 is then used. pThe simulated flow rate Q of the ablation component obtained in step S3.2 m Linear summation is performed at the same time points, and in actual forecasts, the flood rise flow is added to obtain the total forecast runoff Q at the basin outlet at each time point. sim .

[0048] Step S4: Based on the total predicted runoff output in Step S3 and the measured runoff data from the Hero Bridge hydrological station after interpolation processing in Step S1, a multi-objective composite loss function is constructed. Iterative training is then performed using historical flood data to obtain a physical-guided deep learning flood forecasting model. Specifically: Step S4.1: Construct a multi-objective composite loss function. Extract the total predicted runoff Q obtained in step S3.3. sim Compared with the hourly runoff data Q obtained during the flood period from step S1.2 preprocessing t The Kling-Gupta efficiency coefficient (KGE) of both parameters is calculated as the main loss term to evaluate the goodness of fit of the overall process line. At the same time, the extreme value error of both parameters at the flood peak is calculated as the penalty auxiliary term. In addition, hydrophysical boundaries are set for the parameters output by the LSTM neural network model, including three runoff generation coefficients, fast and slow confluence parameters, and a boundary violation penalty term is constructed. The main loss term, penalty auxiliary term, and boundary violation penalty term are weighted to form a multi-objective composite loss function.

[0049] Step S4.2, iterative optimization of the physical-guided deep learning flood forecasting model. The error gradient is calculated using a multi-objective composite loss function, and all learnable parameters in the LSTM neural network model, the fast Nash confluence model, and the slow Nash confluence model are updated using the backpropagation algorithm. Steps 2.4 to 4.1 are repeated until the multi-objective composite loss function reaches its optimum, resulting in a fully calibrated physical-guided deep learning flood forecasting model, abbreviated as PG-LSTM flood forecasting model. Further, in this embodiment, reaching the optimum of the objective composite loss function means minimizing the objective composite loss function value.

[0050] Step S5: Based on the physically guided deep learning flood forecasting model trained in Step S4, input the multi-dimensional feature set of the flood event to be forecasted, and output the full-process flood forecast results and sub-item forecast results. Specifically: Step S5.1: Perform full-process forecast simulation. Acquire multi-source data for the event to be forecasted and calculate a multi-dimensional feature set as in step S1. Input this feature set into the physical-guided deep learning flood forecasting model obtained in step S4. After forward propagation operations by the runoff generation module (step S2) and the runoff confluence module (step S3), directly output the total forecast runoff process line for the entire future time period of the event to be forecasted.

[0051] Step S5.2: Output the sub-project forecast results. Based on the decoupling structure within the physical-guided deep learning flood forecast model, while outputting the total forecast runoff, the runoff for each time period obtained in step S2 and the simulated flow rate Q of the rainfall component output in step S3.1 are extracted separately. p The simulated flow rate Q of the ablation component obtained in step S3.2 m It calculates the proportion of each hydrophysical component in the total runoff and total runoff at each moment, realizing full-process transparency and interpretable forecasting of mixed rain, snow and ice floods in high-altitude cold basins.

[0052] Step S5.3: Render and generate a comparison chart of the PG-LSTM model with superimposed color shadows for flood forecasting with heavy rainfall as the main feature. Figure 2 A comparison chart of the flow process forecast for floods mainly caused by snowmelt and ice melt using a PG-LSTM model that generates overlaid color shadows. Figure 3 A comparison chart of the flood forecast flow process mainly composed of heavy rain and snowmelt / icemelt using a PG-LSTM model that generates overlaid color shadows. Figure 4 ).

[0053] Step S5.4: Render and generate a PG-LSTM model with overlaid color shadows to predict the three runoff processes of rainfall, snowmelt, and ice melt for floods mainly caused by heavy rain. Figure 5 The PG-LSTM model, which generates overlay color shadows, predicts the three runoff processes of rainfall, snowmelt, and icemelt in floods mainly caused by snowmelt and icemelt. Figure 6 The PG-LSTM model, which generates overlay color shadows, predicts the three runoff processes of rainfall, snowmelt, and icemelt in floods mainly caused by a mixture of heavy rain and snowmelt / icemelt. Figure 7 ).

[0054] Step S5.5: Render and generate a PG-LSTM model with overlaid color shadows to predict the confluence of rainfall and snowmelt / icemelt processes for floods dominated by heavy rain. Figure 8 The PG-LSTM model, which generates overlaid color shadows, predicts floods primarily caused by snowmelt and ice melt, using a graph showing the convergence of rainfall and snowmelt / ice melt. Figure 9 The PG-LSTM model, which generates overlaid color shadows, predicts a flood primarily caused by a mixture of heavy rain and snowmelt / icemelt, as shown in the diagram of the two confluence processes of rainfall and snowmelt / icemelt. Figure 10 ).

[0055] The final results obtained from the above steps have good physical significance for exploring the local rain-snow-flood mechanism in the target watershed. The runoff generation coefficient corresponds to the watershed runoff coefficient, and the confluence coefficient can reflect the confluence characteristics of different water sources. The rain-snow-ice mixed flood forecasting method for high-altitude cold watersheds based on physical-guided deep learning has good physical interpretability and provides a more accurate flood forecasting method for the local area.

[0056] In addition, as an optional implementation, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the deep learning flood forecasting method.

[0057] In addition, as an optional implementation, the present invention provides a storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the steps of the deep learning flood forecasting method.

[0058] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for forecasting mixed rain, snow, and ice floods in high-altitude cold watersheds based on physical-guided deep learning, characterized in that... The method for forecasting mixed rain, snow, and ice floods in high-altitude cold basins includes the following steps: Step S1: Obtain multi-source data of the target watershed, preprocess the multi-source data, and form a multi-dimensional feature set based on it. The multi-dimensional feature set includes a static pre-processing feature set and a dynamic time-series feature set. Step S2: Based on the multidimensional feature set of step S1, calculate the effective physical coefficients for rain, snow and ice respectively, construct a long short-term memory neural network model and introduce the effective physical coefficients to map the actual runoff upper limit, and predict the runoff coefficient of each flood through the multidimensional feature set. Multiply the actual runoff upper limit with the runoff coefficient of the flood to obtain the runoff for each time period. Step S3: Based on the runoff generated in each time period obtained in Step S2, two Nash runoff models with different physical response characteristics are constructed for runoff calculation, including the fast Nash runoff model and the slow Nash runoff model. The runoff results of the two are then linearly superimposed to obtain the total predicted runoff at the watershed outlet. Step S4: Based on the total forecast runoff output in step S3 and the measured runoff data from the hydrological station obtained in step S1, a multi-objective composite loss function is constructed, and iterative training is performed using historical flood data to obtain a physical-guided deep learning flood forecasting model. Step S5: Based on the physical-guided deep learning flood forecasting model trained in step S4, input the multi-dimensional feature set of the flood event to be forecasted, and output the full-process forecast results and sub-item forecast results of the flood event.

2. The method for forecasting mixed rain, snow, and ice floods in cold watersheds based on physical-guided deep learning according to claim 1, characterized in that, Specifically, step S1 includes: Step S1.1: Obtain multi-source data for the target watershed, including rainfall data, temperature data, and runoff data; The rainfall data is derived from measured meteorological station data and multi-source fused precipitation datasets, including daily rainfall data, flood season rainfall data, and hourly rainfall data, with a temporal resolution of at least daily. The temperature data is derived from measured meteorological station data and multi-source fused temperature data, including daily average temperature data, daily maximum temperature data, and daily minimum temperature data. The runoff data is derived from measured hydrological station data and multi-source fused runoff data, including daily runoff data and hourly runoff data during flood events. Step S1.2: Preprocess the multi-source data obtained in step S1.1; use linear interpolation to uniformly resample the rainfall and runoff data at different time scales to obtain continuous daily rainfall data P, daily runoff data Q, and hourly rainfall data P during the flood period. t Hourly runoff data Q t The temperature data is resampled using an asymmetric cosine interpolation method to obtain smooth hourly temperature data T. Step S1.3: Based on the data preprocessed in step S1.2, a multidimensional feature set is constructed. The multidimensional feature set includes a dynamic time series feature set and a static early-stage feature set. The static pre-flood feature set is used to characterize the basin background state before a single flood occurs, including the pre-flood snow storage value Sd, the pre-flood rainfall Pa, the pre-flood melting impact SIa, and the previous day's temperature increment ΔT; The dynamic temporal feature set is used to characterize the meteorological driving forces during the flood process, including time-period rainfall, time-period temperature, time-period runoff, and time-period snow storage value divided into each calculation time step.

3. The method for forecasting mixed rain, snow, and ice floods in cold watersheds based on physical-guided deep learning according to claim 2, characterized in that, The static early-stage feature set in step S1.3 includes: The pre-flood snow storage value is calculated daily from the beginning of the year to the day before the flood by setting a temperature threshold and using the degree-day factor method. The pre-flood rainfall is also calculated daily from the beginning of the year to the day before the flood. The pre-flood melting impact is also calculated according to the API model and a decay factor is set according to the characteristics of the watershed. The previous day temperature increment is the temperature difference between the day before the flood and the day the flood begins. The temperature threshold range is consistent with reality.

4. The method for forecasting mixed rain, snow, and ice floods in cold watersheds based on physical-guided deep learning according to claim 3, characterized in that, Specifically, step S2 includes: Step S2.1: Considering the complex mixed flood characteristics of rain, snow, and ice in high-altitude cold basins, based on the time-period rainfall and temperature obtained in Step S1, calculate the effective physical coefficients, including the effective rainfall coefficient P. eff Effective snow coefficient S eff and effective ice coefficient I eff ; Step S2.2: Construct a long short-term memory network model; A Long Short-Term Memory (LSTM) neural network model is constructed as the underlying computational architecture; the LSTM neural network model internally contains memory units c. t And three control gating mechanisms; Step S2.3: Introduce effective physical coefficients to map the actual upper limit of the flow generation; The dynamic temporal feature set obtained in step S1.3 and the effective physical coefficients obtained in step S2.1 are input into the LSTM neural network model constructed in step S2.2 for nonlinear mapping. The model calculates the actual runoff upper limit corresponding to each time step, which represents the physical limits of the three water sources, specifically including: rain runoff upper limit, snow runoff upper limit and ice runoff upper limit. Step S2.4: Predict the runoff coefficient of the flood event using a multidimensional feature set; input the static pre-flood feature set obtained in step S1.3 into the LSTM neural network model to evaluate the initial surface dryness and wetness and water storage status of the entire flood event, and predict the global runoff reduction ratio of the entire flood event, i.e., the runoff coefficient of the flood event, which includes the rain runoff coefficient α, the snow runoff coefficient β and the ice runoff coefficient γ. Step S2.5: Calculate the runoff for each time period; Multiply the actual runoff upper limit of each time period obtained from step S2.3 by the corresponding flood runoff coefficient predicted in step S2.4 to obtain the effective net rainfall generated by each water source at each moment, i.e., the runoff for each time period, including: rain runoff, snow runoff and ice runoff.

5. The method for forecasting mixed rain, snow, and ice floods in cold watersheds based on physical-guided deep learning according to claim 4, characterized in that, In step S2.2: Memory unit c t Used to store long-term hydrological information in sequences; The three control gating mechanisms are input gate i t Forgotten Gate t Output gate o t Input gate i t Controls the proportion of new meteorological information written to memory units at the current moment; forget gate f t Determine how much soil or snow history information from earlier periods to retain; Output gate o t The impact of the state of the control memory unit on the final hydrological output.

6. The method for forecasting mixed rain, snow, and ice floods in cold watersheds based on physical-guided deep learning according to claim 5, characterized in that, The three effective physical coefficients in step S2.1 are used to characterize the maximum runoff potential of rain, snow, and ice in a flood event. The specific calculation method is as follows: The effective rainfall coefficient is the representative coefficient that most significantly affects rainfall runoff. It is calculated by deducting solid snowfall from the temperature threshold range of precipitation phase transition, subtracting losses from vegetation interception and surface depression interception in actual conditions, and considering the infiltration effect under extreme rainstorm conditions. The temperature threshold range of precipitation phase transition is consistent with reality. The effective snow coefficient is the representative coefficient that most significantly affects snowmelt runoff, determined by the current snow storage value and temperature over a given period, and calculated using the degree-day factor method based on temperature. The effective ice coefficient is calculated using the day factor method and is a representative coefficient of the maximum melt flow generated by exposed glaciers under purely thermodynamic driving.

7. The method for forecasting mixed rain, snow, and ice floods in cold watersheds based on physical-guided deep learning according to claim 6, characterized in that, Specifically, step S3 is as follows: Step S3.1: Construct a fast Nash runoff model; input the time-period rainfall runoff calculated in step S2.5 into the fast Nash runoff model to calculate the simulated flow rate of the rainfall component at the watershed outlet, and obtain the simulated flow rate Q of the rainfall component. p : （1） in, This represents the number of linear reservoirs connected in series in the fast Nash confluence model; For the water storage calculation constant of a single reservoir in the fast Nash confluence model; It is a gamma function; For which time period; Step S3.2: Construct a slow Nash confluence model; simultaneously input the time-period snow runoff and time-period ice runoff calculated in step S2.5 into the slow Nash confluence model to calculate the simulated flow rate of the ablation component at the watershed outlet, and obtain the simulated flow rate Q of the ablation component. m : （2） in, This represents the number of linear reservoirs connected in series in the slow Nash confluence model; K m For the water storage calculation constant of a single reservoir in the slow Nash confluence model; Step S3.3, calculate the total predicted runoff; simulate the flow rate Q of the rainfall component obtained in step S3.

1. p The simulated flow rate Q of the ablation component obtained in step S3.2 m Linear summation is performed at the same time points, and in actual forecasts, the flood rise flow is added to obtain the total forecast runoff at the basin outlet at each time point. .

8. The method for forecasting mixed rain, snow, and ice floods in cold watersheds based on physical-guided deep learning according to claim 7, characterized in that, Specifically, step S4 includes: Step S4.1: Construct a multi-objective composite loss function; extract the total predicted runoff Q obtained in step S3.

3. sim Compared with the hourly runoff data Q obtained from the preprocessing in step S1.2 during the flood period t The Kling-Gupta efficiency coefficient (KGE) of both models is calculated as the main loss term. Simultaneously, the extreme value error of both models at the peak flood time is calculated as the penalty auxiliary term. In addition, hydrophysical boundaries are set for the parameters output by the LSTM neural network model, and an out-of-bounds penalty term is constructed. The main loss term, penalty auxiliary term, and out-of-bounds penalty term are weighted to form a multi-objective composite loss function. Step S4.2: Iterative optimization of the physical-guided deep learning flood forecasting model; calculate the error gradient using the multi-objective composite loss function, and update all learnable parameters in the LSTM neural network model, the fast Nash confluence model, and the slow Nash confluence model through the backpropagation algorithm; repeat steps 2.4 to 4.1 until the multi-objective composite loss function finds its optimal value, and obtain the parameter-calibrated physical-guided deep learning flood forecasting model, abbreviated as PG-LSTM flood forecasting model.

9. The method for forecasting mixed rain, snow, and ice floods in cold watersheds based on physical-guided deep learning according to claim 8, characterized in that, In step S4: In step S4.1, the parameters output by the LSTM neural network model include three flow generation coefficients, fast and slow flow parameters; In step S4.2, the condition for finding the optimal multi-objective composite loss function is that the multi-objective composite loss function value converges to the minimum value.

10. The method for forecasting mixed rain, snow, and ice floods in cold watersheds based on physical-guided deep learning according to claim 9, characterized in that, Specifically, step S5 includes: Step S5.1: Perform full-process forecast simulation; acquire meteorological forecast data for the event to be forecasted and calculate a multi-dimensional feature set according to step S1; input it into the physical-guided deep learning flood forecasting model obtained in step S4; after steps S2 and S3, output the total forecast runoff process line for the entire future time period of the event to be forecasted. Step S5.2: Output the sub-project forecast results; relying on the decoupling structure within the physical-guided deep learning flood forecast model, while outputting the total forecast runoff, extract the runoff for each time period obtained in step S2 and the simulated flow rate Q of the rainfall component output in step S3.

1. p The simulated flow rate Q of the ablation component obtained in step S3.2 m It calculates the proportion of each hydrophysical component in the total runoff and total runoff at each moment, realizing full-process transparency and interpretable forecasting of mixed rain, snow and ice floods in high-altitude cold basins.

Images