Deep learning method for runoff simulation in snowmelt-rain mixed recharge basin

Abstract

This invention discloses a deep learning-based runoff simulation method for mixed rain-snowfall-snowmelt watersheds, comprising: acquiring meteorological driving data of the target watershed, inputting it into a snowmelt dynamic mechanism model for transformation to obtain effective liquid water; extracting physical state variables using an effective liquid water-driven hydrological mechanism model, and constructing a multi-dimensional fusion feature system accordingly; performing pre-screening on the sequence of the multi-dimensional fusion feature system based on preset physical consistency conditions to obtain candidate feature trajectories that satisfy objective hydrological evolution laws; constructing ensemble features based on each candidate feature trajectory, inputting the ensemble features into a pre-trained time-series mapping network, and outputting simulated runoff results. This invention effectively integrates the boundary constraints of mechanism deduction with the fitting ability of deep learning, improving the accuracy, stability, and physical interpretability of runoff simulation in mixed-supply watersheds in high-altitude cold regions during snowmelt sensitive periods.

Description

Technical Field

[0001] This invention relates to the field of hydrological simulation and forecasting technology, and in particular to a deep learning-based runoff simulation method for mixed rain-snow-melt watersheds. Background Technology

[0002] High-altitude cold-region watersheds are the source of many major rivers, and their runoff formation process is controlled by a combination of rainfall, snow accumulation, and snowmelt release. Accurate runoff simulation is of great technical significance for revealing the water cycle mechanism under complex underlying surface conditions and understanding the spatiotemporal distribution patterns of seasonal water resources. Especially in the context of climate change, establishing high-precision runoff physical extrapolation and quantitative analysis methods is the core technical foundation for ensuring the safe allocation of water resources in watersheds.

[0003] Currently, runoff simulation for such watersheds mainly relies on traditional mechanistic hydrological models or purely data-driven deep learning models. Traditional mechanistic models depend on extensive empirical parameter calibration, and when depicting hydrological phase transitions and deep physical storage processes, they often face limitations such as structural rigidity and the similarity of different parameters with the same effect, leading to larger simulation errors during the spring flow rise phase. Deep learning models, which have been widely used in recent years, typically establish a nonlinear mapping between initial meteorological factors and runoff. This data-driven paradigm, when dealing with long-sequence evolution, is easily limited by the quality of training samples and random parameter noise. When facing complex mixed conditions with severe time delays caused by snowfall and snowmelt, it often exhibits specific problems such as process fitting distortion, physical boundary failure, and a decline in the ability to capture extreme peaks.

[0004] In summary, existing methods struggle to effectively balance simulation accuracy with the objective rationality of physical processes when dealing with complex, multi-source hydrological processes in high-altitude and cold regions. Furthermore, their ability to generalize data feature extraction during seasonal recharge transitions remains insufficient. Therefore, it is necessary to investigate a runoff simulation method that enhances the robustness and accuracy of data processing in complex mixed recharge scenarios, thereby overcoming the common technical challenges faced by existing technologies in long-term hydrological forecasting. Summary of the Invention

[0005] Therefore, this application provides a deep learning-based runoff simulation method for rain-snow-melt mixed recharge watersheds, aiming to solve the aforementioned problems existing in the prior art.

[0006] A deep learning-based runoff simulation method for mixed rain-snowfall-melt watersheds includes:

[0007] Acquire meteorological driving data for the target watershed;

[0008] Meteorological driving data are input into the snowmelt dynamic mechanism model to obtain effective liquid water;

[0009] Physical state variables are extracted using an effective liquid water-driven hydrological mechanism model, and a multi-dimensional fusion feature system is constructed accordingly.

[0010] Based on preset physical consistency conditions, the sequences of the multidimensional fusion feature system are screened to obtain candidate feature trajectories;

[0011] A set of features is constructed based on each candidate feature trajectory. The set of features is then input into a pre-trained time-series mapping network to output simulated runoff results.

[0012] Before inputting the set features into the pre-trained temporal mapping network, there is also a pre-training process for offline construction of the deep learning model:

[0013] Extract the values ​​of multiple candidate feature trajectories on the same physical state variable at each time step and calculate the mean;

[0014] Calculate the sum of squares of the differences between the corresponding variable values ​​and the mean of each candidate feature trajectory, and extract the arithmetic square root of the sum of squares after dividing by the total number of trajectories to obtain the standard deviation characterizing the dispersion of the multi-trajectory distribution;

[0015] The mean and standard deviation are concatenated and fused to form a set feature sequence, and the set feature sequence is truncated according to the set sliding time window scale to construct supervised training samples;

[0016] The weight parameters of the initial deep learning network are iteratively updated using supervised training samples, thereby obtaining a pre-trained temporal mapping network.

[0017] The above technical solutions can effectively integrate the boundary constraints of mechanism deduction with the fitting ability of deep learning, thereby improving the accuracy, stability and physical interpretability of runoff simulation in mixed-supply basins in high-altitude and cold regions during the snowmelt sensitive period. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 This is a schematic diagram of a deep learning-based runoff simulation method for mixed rain-snow-melt watersheds provided in this application.

[0020] Figure 2 This is a schematic diagram of the snow-constrained snow melting process provided in this application.

[0021] Figure 3 This is a schematic diagram of the liquid water confined refreezing process provided in this application.

[0022] Figure 4This is a schematic diagram of the pre-training process for offline construction of deep learning models provided in this application.

[0023] Figure 5 This is a schematic diagram of the verification logic flow for screening using long-term water balance constraints provided in this application.

[0024] Figure 6 This is a schematic diagram of another deep learning runoff simulation method for mixed rain-snow-melt watersheds provided in this application.

[0025] Figure 7 This is a simulation diagram of daily runoff for selected years provided in this application. Detailed Implementation

[0026] For ease of understanding, some symbols in this specification are defined as follows. Unless otherwise stated, the following symbols have the following meanings:

[0027] SWE temp,t SWE t * Both represent the intermediate snowmelt equivalent at the current moment;

[0028] SWE prev SWE t-1 Both represent the pre-initialized snow water equivalent of the previous moment;

[0029] M pot,t M t pot Both represent the potential snowmelt volume at the current moment;

[0030] RF pot,t RF t pot Both represent the potential refreeze amount at the current moment.

[0031] Example 1: This example provides a deep learning-based runoff simulation method for rain-snow-melt mixed recharge watersheds, such as... Figure 1 As shown, the method may include the following steps:

[0032] Step 101: Obtain meteorological driving data for the target watershed.

[0033] Furthermore, meteorological driving data are specifically obtained through collection from meteorological observation stations or extraction from reanalysis datasets. The high-altitude, cold region where the target watershed is located presents unique climatic conditions, making it difficult for conventional runoff simulations to accurately depict its complex hydrological cycle. The acquired meteorological sequences encompass all the fundamental elements driving hydrological changes within the watershed, constituting the source input for the entire model calculation chain. In actual operation, the acquired meteorological driving data needs to undergo time benchmark unification and outlier standardization. This processing can employ extreme value normalization algorithms to ensure scale consistency across data of different magnitudes.

[0034] Step 102: Input the meteorological driving data into the snowmelt dynamic mechanism model to obtain effective liquid water.

[0035] Specifically, the snowmelt dynamics model is a computational module used to handle the physical processes of solid-liquid precipitation phase transition, snow accumulation, and ablation. In high-altitude watersheds, runoff recharge is controlled by both rainfall and snowmelt, making it difficult for a single precipitation input to directly drive the runoff generation model. After inputting meteorological driving data into the model, it performs phase separation of precipitation based on temperature characteristics and, combined with environmental energy conditions, extrapolates the physical actions of ablation and refreezing within the snow. Through this physical extrapolation, the model outputs effective liquid water that satisfies the mass conservation constraint. Effective liquid water integrates the volume of pure liquid precipitation and liquid water released from the melting of solid snow, eliminating overlapping and loss components, and providing a physically accurate water input benchmark for subsequent runoff generation calculations.

[0036] Optionally, the various physical constraint parameters within the snowmelt dynamic mechanism model can be sampled and assigned values ​​through a preset uniformly distributed parameter space, thereby adapting to the differences in snow and ice dynamics at different altitudes.

[0037] Step 103: Extract physical state variables using an effective liquid water-driven hydrological mechanism model, and construct a multi-dimensional fusion feature system based on these (physical state variables).

[0038] Specifically, this can be further implemented by using effective liquid water to drive hydrological mechanism models based on multiple sets of candidate parameter configurations to extract physical state variables, and thereby constructing multiple sets of multidimensional fusion feature system sequences.

[0039] For example, the hydrological mechanism model can be implemented using the three-source Xin'anjiang model. After the effective liquid water, acting as the source driving water volume, enters the system, the model internally performs a pipeline extrapolation based on physical laws such as evaporation conversion, runoff generation upon saturation, and water source delineation. During this extrapolation process, the system simultaneously extracts various physical state variables reflecting the internal evolution of the watershed. Since purely data-driven models suffer from black-box limitations and struggle to perceive the water distribution within the watershed, extracting these physical state variables allows for the structured encapsulation of external meteorological drivers, intermediate transition processes, surface and groundwater storage states, and the final response processes of each source. The resulting multi-dimensional fusion feature system constitutes a feature set covering the complete causal chain, providing physically constrained guidance information for subsequent nonlinear mapping networks.

[0040] In actual implementation, the system is based on multiple sets of candidate parameters for hydrological mechanism models generated in advance through a random sampling strategy (see Example 6 for details on the generation method), and configures the above-mentioned hydrological mechanism models and drives them to run one by one, thereby obtaining multiple sets of corresponding multidimensional fusion feature system time series.

[0041] Step 104: Based on the preset physical consistency conditions, the sequences of each group of multidimensional fusion feature systems are screened to obtain candidate feature trajectories.

[0042] The multidimensional fusion feature system is a time series set generated from multiple randomly sampled model parameters. Since random parameter combinations may cause the model to deviate from objective physical laws, multidimensional quantitative physical consistency conditions must be introduced as a pre-screening criterion. These physical consistency conditions specifically include digital thresholds for system closure residuals, physical value boundaries of variables, and logical relationships among various runoff components. Performing consistency checks on each generated feature trajectory sequence automatically eliminates abnormal sequences containing non-physical negative values, water imbalances, or seasonal runoff logical distortions. The remaining sequences are defined as candidate feature trajectories. This action improves the quality of the dataset and eliminates potentially contaminated physical fallacies that could pollute subsequent deep feature learning.

[0043] In some alternative implementations, when performing the filtering action, physical consistency conditions of different dimensions can be set as serial filtering nodes, and only sequences that simultaneously satisfy all threshold constraints can be adopted as candidate feature trajectories.

[0044] Step 105: Construct a set of features based on each candidate feature trajectory, input the set of features into a pre-trained time-series mapping network, and output the simulated runoff results.

[0045] Further implementation involves: statistically aggregating each candidate feature trajectory to construct a set of features, inputting the set of features into a pre-trained time-series mapping network, and outputting simulated runoff results.

[0046] In this embodiment, the construction of ensemble features can be achieved by performing statistical aggregation calculations on multiple selected candidate feature trajectories at horizontal time steps. Since a single physical trajectory is insufficient to fully characterize the parameter uncertainties of complex hydrological systems in high-altitude and cold regions, the central trend and discrete distribution indicators of each trajectory are extracted and fused together. The resulting ensemble features thus possess parameter uncertainty encoding information. The pre-trained time-series mapping network can adopt a long short-term memory network architecture, and its internal network weights have been iteratively optimized using historical sample sets. After the ensemble features are input into this network, the network utilizes its internal gating mechanism to extract time-series dependencies, learns the nonlinear mapping logic between multi-level physical process variables and outlet runoff, thereby generating and outputting simulated runoff results for the target time period. This result combines the physical causality of the mechanistic model with the generalization fitting ability of deep learning, avoiding the flow rise fitting error caused by parameter simplification during the snowmelt sensitive period.

[0047] Example 2: Based on Example 1 above, the steps for inputting meteorological driving data into the snowmelt dynamic mechanism model to obtain effective liquid water are further described in detail.

[0048] In one possible implementation, the meteorological driving data includes the total precipitation and average temperature at the current moment; the step of inputting the meteorological driving data into the snowmelt dynamic mechanism model to obtain effective liquid water specifically includes a precipitation phase classification process:

[0049] Based on average air temperature and a preset pure snow temperature threshold T s and the preset pure rain temperature threshold T r The numerical relationship between the two phases is used to classify the total precipitation into solid precipitation and liquid precipitation.

[0050] Furthermore, the precipitation phase classification process specifically involves real-time monitoring of the atmospheric thermodynamic state in high-altitude and cold regions to determine the physical and geometric morphology of the total water falling to the surface. In open, high-altitude watersheds, the runoff generation mechanisms of solid and liquid precipitation exhibit significant heterogeneity; accurate identification of phase distribution is a prerequisite for constructing high-precision hydrologically driven flow. In practice, the determination process is based on comparing the current average temperature sequence with two preset critical phase transition temperatures.

[0051] When the average temperature T t Satisfy T t ≤T s When the calculation relationship is as follows:

[0052] P snow,t =P t Among them, P snow,t Let P be the solid precipitation at time t. tLet t be the total precipitation at time t.

[0053] P rain,t =0; where P rain,t Let be the amount of liquid precipitation at time t.

[0054] When the average temperature satisfies T t ≥T r When the calculation relationship is as follows:

[0055] P rain,t =P t Among them, P rain,t Let P be the amount of liquid precipitation at time t. t Let t be the total precipitation at time t.

[0056] P snow,t =0; where P snow,t Let be the solid precipitation at time t.

[0057] When the average temperature satisfies T s <T t <T r When the calculation relationship is as follows:

[0058] P rain,t =P t *(T t -T s ) / (T r -T s );

[0059] Among them, P rain,t Let P be the amount of liquid precipitation at time t. t Let T be the total precipitation at time t. t Let T be the average temperature at time t. s T is the preset pure snow temperature threshold. r This is the preset pure rain temperature threshold.

[0060] P snow,t =P t -P rain,t ;

[0061] Among them, P snow,t Let P be the solid precipitation at time t. t Let P be the total precipitation at time t. rain,t Let be the amount of liquid precipitation at time t.

[0062] As an example, a simplified dimensionless calculation can be used to illustrate the above phase division logic. Assume that after standardization, the total precipitation at the current moment is 10.0, the average temperature is 0.5, the preset pure snow temperature threshold is set to 0.0, and the preset pure rain temperature threshold is set to 1.0. Since the average temperature at the current moment is between the pure snow temperature threshold and the pure rain temperature threshold, i.e., satisfying the numerical relationship 0.0 < 0.5 < 1.0, the system triggers linear interpolation calculation. Substituting into the formula, the liquid precipitation is calculated as: 10.0 * (0.5 - 0.0) / (1.0 - 0.0) = 5.0; and the solid precipitation is calculated as: 10.0 - 5.0 = 5.0. The above dimensionless calculation shows that the model can smoothly achieve a continuous transition of precipitation phases.

[0063] In some optional implementations, the preset pure snow temperature threshold can be calibrated within the range of -2.5℃ to 0.5℃ based on historical observation data of the watershed, and the preset pure rainfall temperature threshold can be calibrated within the range of 0.5℃ to 3.0℃, thereby more accurately reflecting the microclimate characteristics of the predetermined plateau watershed. The above parameter ranges are only recommended prior values ​​based on observation experience in typical cold regions, and can be adjusted adaptively according to watershed characteristics and simulation results in practical applications. In addition, nonlinear exponential or polynomial transition curves can also be used as alternatives for phase division relationships to depict more complex phase mixing states.

[0064] One possible implementation method, after obtaining solid precipitation and liquid precipitation, includes: using a pre-configured snowfall correction coefficient to numerically compensate the solid precipitation to obtain the corrected solid precipitation, so as to eliminate the measurement deviation of solid precipitation in the process of on-site meteorological observation; the corrected solid precipitation replaces the solid precipitation in subsequent calculations.

[0065] Specifically, in high-altitude and cold regions, traditional instrument-measured snowfall processes are easily affected by wind-induced runoff caused by strong near-surface winds, resulting in instrument-measured solid precipitation values ​​that are usually significantly lower than the actual snowfall received by the surface. To address this technical bottleneck, this step introduces a numerical compensation process. The solid precipitation values ​​obtained in the previous steps are multiplied by a pre-configured snowfall correction coefficient to reconstruct the water volume distribution.

[0066] P snow_corr,t =P snow,t *SFCF;

[0067] Among them, P snow_corr,t Let P be the corrected solid precipitation at time t. snow,t Let be the solid precipitation at time t, and SFCF be the pre-configured snowfall correction factor.

[0068] For example, suppose the current solid precipitation value is 5.0, and the pre-configured snowfall correction factor is set to 1.10. Substituting into the above multiplication relationship for numerical compensation calculation: 5.0 * 1.10 = 5.5. Thus, the corrected solid precipitation value is 5.5, effectively compensating for the instrument capture loss deviation caused by wind speed disturbance.

[0069] In some optional implementations, the pre-configured snowfall correction coefficient (SFCF) is typically optimized within the range of 0.8 to 1.2. It should be noted that this range is only a priori reference based on typical high-altitude and cold-climate stations, and the actual value can be flexibly determined according to watershed characteristics and verification results. Optionally, this coefficient can also be established as a dynamic function of near-surface wind speed and air temperature to achieve adaptive daily correction of measurement deviations throughout the entire time period.

[0070] In some alternative embodiments, refer to Figure 2 The step of inputting meteorological driving data into the snowmelt dynamic mechanism model to obtain effective liquid water also includes the snow accumulation-constrained snowmelt process:

[0071] The intermediate snow water equivalent is obtained by updating the pre-initialized snow water equivalent of the previous moment based on solid precipitation.

[0072] The potential snowmelt amount is calculated based on the average air temperature, the pre-configured snowmelt initiation temperature, and the pre-configured degree-day factor.

[0073] Using the intermediate snow water equivalent as the upper limit, the potential snow melt amount is calculated with a minimum constraint to obtain the actual snow melt amount.

[0074] Furthermore, the snow accumulation constrained snowmelt process is used to precisely characterize the water ablation trajectory within the surface snow layer. If the ablation calculation does not consider the actual existing snow reservoir capacity on the ground, it can easily lead to false snowmelt outflow in the absence of snow on the surface, resulting in a serious distortion of the total spring runoff. This step first overlays the existing snow reservoir status, merging the historically accumulated snow water equivalent with the newly added solid precipitation of the day, to establish the boundary of the intermediate snow water equivalent available for consumption at the current moment.

[0075] SWE temp,t =SWE prev +P snow,t ;

[0076] Among them, SWE temp,t SWE is the current intermediate snowmelt equivalent. prev P is the pre-initialized equivalent of snowmelt from the previous time step. snow,t This refers to solid precipitation.

[0077] In this step, if the snowfall correction coefficient compensation has been performed in the previous step, the solid precipitation refers to the corrected value after numerical compensation.

[0078] M pot,t =DDF*max(T t -T0,0);

[0079] Among them, M pot,t T represents the potential snowmelt at the current moment, DDF is the day factor, and T is the day factor. t T0 is the average air temperature, T0 is the pre-configured snow melting start temperature, and max() is the maximum selection operator.

[0080] M t =min(M pot,t SWE temp,t );

[0081] Among them, M t M represents the actual snowmelt amount at the current moment. pot,t SWE represents the potential snowmelt volume at the current moment. temp,t Let be the intermediate snow water equivalent at the current moment, and min() be the minimum constraint operator.

[0082] In this embodiment, the degree-day factor (DDF) is set to 4.0 mm / (℃·d), and the pre-configured snowmelt initiation temperature (T0) is set to 0.5℃. In practice, the values ​​can be adjusted adaptively according to the watershed radiation characteristics and snow accumulation evolution process.

[0083] For example, when the average daily temperature is higher than the initial snowmelt temperature, the potential snowmelt volume is calculated and constrained by the upper limit of the snow water equivalent. In this case, because the temperature is higher than the initial temperature, the temperature difference term in the refreezing formula is zero, and the actual refreezing volume is zero. The effective liquid water is the sum of liquid precipitation and actual snowmelt volume, and the snow water equivalent is reduced accordingly.

[0084] In addition, as an alternative, ablation calculations can also be performed by combining radiation terms and aerodynamic terms based on the principle of energy balance.

[0085] Optional, refer to Figure 3 The process of inputting meteorological driving data into the snowmelt dynamic mechanism model to obtain effective liquid water also includes a liquid water-constrained refreezing process:

[0086] The potential refreezing amount is calculated based on the pre-configured snow melting initiation temperature, average air temperature, pre-configured refreezing coefficient, and degree-day factor.

[0087] Based on the amount of liquid precipitation and the pre-configured snow layer liquid water interception ratio coefficient, the upper limit of liquid water constraint is determined;

[0088] Using the upper limit of liquid water constraint as the standard, the potential refreezing amount is calculated by minimizing the constraint to obtain the actual refreezing amount.

[0089] Specifically, the liquid water-constrained refreezing process describes the physical phenomenon of free liquid water within the snow layer hardening and refreezing when the ambient temperature drops below the ablation point. Based on daily computational efficiency, it is assumed that the amount of liquid water retained within the snow layer is limited, and the primary constraint comes from the daily increase in precipitation. Therefore, in calculating the potential refreezing capacity driven by low temperatures over the same period, the actual available free liquid precipitation within the snow layer must be used as a physical upper limit to prevent false over-refreezing.

[0090] RF pot,t =C rf *DDF*max(T0-T t ,0);

[0091] Among them, RF pot,t C represents the potential refreeze amount at the current moment. rf Here, DDF is the pre-configured refreezing coefficient, T0 is the degree-day factor, and T is the snow melt initiation temperature. t Let be the average temperature, and max() be the maximum selection operator.

[0092] RF lim,t =κ*P rain,t ;

[0093] Among them, RF lim,t κ represents the upper limit of liquid water constraint at the current moment, κ is the pre-configured liquid water retention ratio coefficient in the snow layer, and P is the upper limit of liquid water constraint at the current moment. rain,t This refers to liquid precipitation.

[0094] RF t =min(RF pot,t ,RF lim,t );

[0095] Among them, RF t RF represents the actual refreeze amount at the current moment. pot,t RF represents the potential refreeze amount at the current moment. lim,t is the upper limit of the liquid water constraint at the current moment, and min() is the minimum constraint operator.

[0096] In this embodiment, the pre-configured refreezing coefficient C rf The value is set to 0.05, and the snow layer liquid water interception ratio coefficient κ is set to 0.08. In practice, it can be adaptively adjusted according to the snow layer structure and freeze-thaw cycle characteristics.

[0097] As an example, when the average daily temperature is below the snowmelt initiation temperature, the potential snowmelt is zero. The potential refreezing is calculated and capped by an upper limit constrained by liquid water. Effective liquid water equals liquid precipitation minus actual refreezing. The snowmelt equivalent increases due to refreezing replenishment.

[0098] In addition, an alternative approach could be to incorporate the residual water holding capacity of the snow layer that did not flow out the previous day for cumulative calculation, in order to construct a more refined evolution path of water transport within multi-layered snow cover.

[0099] Optionally, the step of inputting meteorological driving data into the snowmelt dynamic mechanism model to obtain effective liquid water also includes a water conservation update process:

[0100] Based on the actual snow melt volume and the actual refreezing volume, the intermediate snow water equivalent is dynamically corrected to obtain the snow water equivalent at the current moment.

[0101] The effective liquid water is generated by adding the liquid precipitation to the actual snow melt and subtracting the actual refreezing.

[0102] Among them, the sum of the dynamic changes in effective liquid water and snow water equivalent equals the total precipitation.

[0103] Furthermore, the water conservation update process ensures the mass closure of the entire snowmelt dynamic mechanism model. By subtracting the ablation term and adding the refreezing term, the intermediate snow water equivalent is smoothly backfilled and updated to the final solid snow reservoir storage state at the current moment. In the same period, the liquid precipitation flowing on the surface is combined with the actual snowmelt released by the snow accumulation, and the water consumed by refreezing is deducted to reconstruct the effective liquid water that is ultimately used to drive the runoff, thus establishing a strict water conservation chain.

[0104] SWE t =SWE temp,t -M t +RF t ;

[0105] Among them, SWE t SWE is the equivalent of snowmelt at the current moment. temp,t M is the equivalent of intermediate snowmelt. t RF represents the actual amount of snow melted. t This represents the actual amount of refreezing.

[0106] P eff,t =P rain,t +M t -RF t ;

[0107] Among them, P eff,t For the effective liquid water generated, P rain,t M represents liquid precipitation. t RF represents the actual amount of snow melted. t This represents the actual amount of refreezing.

[0108] P eff,t +SWE t -SWE prev =P t ;

[0109] Among them, P eff,t To generate effective liquid water, SWE t SWE is the equivalent of snowmelt at the current moment. prev P is the pre-initialized equivalent of snowmelt from the previous time step. t This represents the total precipitation.

[0110] It should be noted that when no snowfall correction factor is introduced, i.e., SFCF=1.0 or no numerical compensation is performed, the total precipitation in the above water conservation equation is the original observed total precipitation P. t When a snowfall correction factor is introduced, the total precipitation should be understood as the total amount of water that actually enters the system after correction, i.e., the sum of liquid precipitation and corrected solid precipitation.

[0111] The physical meaning of the aforementioned water conservation relationship is that, within any single time step, the total water input entering the system is completely distributed across two channels: effective liquid water production and changes in solid snow inventory. There is no passive increase or spurious consumption. When the ambient temperature is higher than the snowmelt initiation temperature, the system generates snowmelt while refreezing is zero, and the effective liquid water consists of the sum of liquid precipitation and snowmelt. When the ambient temperature is lower than the snowmelt initiation temperature, the system generates refreezing while snowmelt is zero, and the effective liquid water consists of liquid precipitation minus refreezing. In both cases, substituting into the formula verifies that the sum of the dynamic changes in effective liquid water and snow water equivalent is strictly equal to the total water input entering the system, thus achieving global quality closure. This multi-variable linkage closed loop suppresses the phenomena of spurious water generation and loss in the calculation process.

[0112] In some alternative implementations, the aforementioned water conservation update process can serve as a hydrological mechanism check operator, monitoring the quality of the data stream calculation in real time and automatically triggering a warning when a residual non-zero value is detected. Alternatively, a direct snow sublimation loss term can be introduced, incorporating the sublimation amount as a separate deduction branch into the balance equation.

[0113] Example 3: Based on Example 1 above, the process of extracting physical state variables and constructing a multi-dimensional fusion feature system is described in further detail.

[0114] In one possible implementation, during the step of acquiring meteorological driving data of the target watershed, the potential evapotranspiration of the target watershed is also acquired simultaneously; during the step of extracting physical state variables and constructing a multi-dimensional fusion feature system based on them:

[0115] The physical state variables are specifically extracted as follows: snow melt, snow storage (snow water equivalent), soil moisture content, simulated total runoff, surface runoff, soil flow, and groundwater runoff;

[0116] The multidimensional fusion feature system consists of eleven variables at four physical levels, specifically including: meteorological driving data and pre-acquired potential evapotranspiration representing external driving forces; effective liquid water, snowmelt volume and snow storage representing snowmelt transition; soil moisture content representing water storage state; and simulated total runoff, surface runoff, soil flow and groundwater runoff representing source responses.

[0117] Furthermore, the multi-dimensional fusion feature system is specifically implemented by temporally splicing and recombining the internal implicit states generated by the hydrological mechanism model with the external meteorological boundary conditions. Conventional deep learning algorithms suffer from black-box defects when processing hydrological time series predictions, making it difficult to capture the spatial distribution and physical transformation state of water volume within the watershed. Extracting physical state variables to construct a feature system covering the complete causal chain can provide prior information with physical boundary constraints for subsequent mapping networks. In specific execution, effective liquid water is used as net rainfall input to drive the three-source Xin'anjiang model to perform runoff calculations, and the intermediate state outputs of each simulation node are captured and recorded simultaneously.

[0118] The first physical level of the aforementioned multidimensional fusion feature system is the level that characterizes external drivers. It consists of the total precipitation and average temperature contained in the meteorological driving data, together with the pre-acquired potential evapotranspiration, providing a total of 3 independent feature dimensions. This level defines the external boundary of the target watershed system that receives atmospheric energy and moisture input.

[0119] The second physical level of the aforementioned multidimensional fusion feature system is the level that characterizes the snow melting transition, which includes effective liquid water, snow melting amount and snow storage, providing a total of 3 feature dimensions. This level records the mass retention delay and secondary release trajectory of water during the solid-liquid phase transformation process.

[0120] The third physical level is the level that characterizes the water storage state, specifically the soil moisture content, providing a total of 1 feature dimension. This level characterizes the pre-dry and wet conditions of the watershed underlying surface and the current storage capacity of the vadose zone reservoir.

[0121] The fourth physical level is the level that characterizes the source response, which includes simulated total runoff, surface runoff, soil runoff and groundwater runoff, providing a total of 4 characteristic dimensions, reflecting the water receding dynamics process of effective liquid water forming fast and slow flow components after being regulated by different media in the watershed.

[0122] To adapt to the requirements of computer processing and the input interface of deep learning networks, the eleven-dimensional variables extracted from the four physical layers are structurally concatenated into a temporal feature matrix in physical memory. Assuming the sliding time window length is L, for any prediction time, the eleven-dimensional variables are truncated from the historical sequence for L consecutive time steps and concatenated along the column dimension to form a two-dimensional data tensor of size (L, 11). This data tensor structure solidifies the entire process of rain-snow conversion, water infiltration, and multi-source confluence evolution in the high-altitude cold region's rain-snow mixed recharge basin within the channel sequence, defining the physical boundaries for subsequent network optimization.

[0123] In some alternative implementations, the hydrological mechanism model for extracting physical state variables is not limited to the Xin'anjiang model with three water sources. It can also be a hydrophysical model with a clear runoff generation and diversion mechanism, such as HBV (a conceptual watershed hydrological model) or SWAT (soil and water assessment tool).

[0124] In another alternative implementation, the hydrological mechanism model is implemented using the HBV model. The HBV model uses available liquid water as the net precipitation input, calculates actual evaporation and soil water storage via a soil moisture calculation routine, and generates fast and slow flow components respectively through the linear drainage relationship between the upper and lower groundwater tanks, ultimately merging to output simulated runoff. The extracted physical state variables include soil moisture content, upper groundwater storage, lower groundwater storage, and the corresponding fast and slow flow components.

[0125] Furthermore, for extremely cold basins that include permanent glacier recharge, the multidimensional fusion feature system can expand upon the above four physical levels by adding a physical level that characterizes the glacier ablation state, and incorporate the extracted bottom-level glacier meltwater runoff as an independent feature dimension into the column dimension of the time-series feature matrix, thereby expanding the basin applicability boundary of the feature system.

[0126] Example 4: Based on Example 3 above, this example further describes the detailed process of filtering the sequence of the multidimensional fusion feature system based on preset physical consistency conditions to obtain candidate feature trajectories.

[0127] In one possible implementation, the physical consistency condition comprises multiple combined constraints on the system's closed state, the physical boundaries of variables, and the logic of generation, specifically including:

[0128] Long-term water balance constraints for controlling the closure residuals of water volume in multi-year systems;

[0129] Nonnegativity constraints that define the lower bounds of each state variable in a multidimensional fusion feature system;

[0130] Seasonal snowmelt phase constraints based on pre-defined natural month intervals to check the peak flow periods;

[0131] Source runoff conservation constraints that limit the deviation between the sum of multiple water source runoff components and the total flow;

[0132] And a performance threshold constraint that controls the lowest limit of historical fit based on pre-acquired historical measured runoff data.

[0133] Specifically, while the massive parameter set generated by random sampling based on hydrophysical mechanism models expands the candidate sample space, random parameter configuration is prone to producing abnormal characteristic trajectories that violate objective physical evolution laws. The multidimensional quantitative physical consistency conditions constituted by the above five constraints can effectively construct a pre-screening gate for features.

[0134] Specifically, the non-negativity constraint is achieved by ensuring that the values ​​of snow storage, soil moisture content, and each runoff component in the multi-dimensional fusion feature system are all ≥0 over the entire time series, thus eliminating invalid trajectories that generate negative water values. The seasonal snowmelt phase constraint is achieved by monitoring the month of peak flow in the high-altitude cold regions of the Northern Hemisphere. The system automatically compares and eliminates sequences whose peak months do not fall within a preset interval of common snowmelt periods in nature. This preset interval can be configured to be from March to July to ensure that the snowmelt features extracted by the model conform to seasonal climate rhythms. The performance threshold constraint is achieved by calculating the historical goodness of fit during the training period, eliminating inefficient parameter groups with Nash efficiency coefficients lower than the preset baseline requirement. The specific value of the preset baseline requirement can be adaptively set by those skilled in the art based on the complexity of the hydrological characteristics of the target watershed and the required retention ratio of candidate trajectories; typically, a value within the range of 0.2 to 0.5 is sufficient to meet the screening requirements. This ensures that the prior samples for subsequent mapping network learning have sufficient information content.

[0135] One possible implementation is, such as Figure 5 As shown, the verification logic for screening using long-term water balance constraints specifically includes:

[0136] Obtain the pre-estimated actual evaporation rate for the corresponding time period;

[0137] The total effective liquid water input to the system over multiple time scales is statistically analyzed, and the actual evapotranspiration, simulated total runoff, and watershed water storage changes calculated from soil moisture content are subtracted from the total to obtain the system closure residual.

[0138] The system calculates the relative ratio between the system closure residual and the total amount of effective liquid water, and determines that the corresponding characteristic trajectory passes the long-term water balance constraint when the relative ratio is less than or equal to the preset balance threshold.

[0139] Actual evapotranspiration can be obtained from the internal evaporation conversion module of the hydrological mechanism model. This module calculates the actual evapotranspiration value based on the input potential evapotranspiration and the current soil moisture state, according to the constraint relationship between evaporation capacity and water supply conditions. Those skilled in the art can select the corresponding actual evapotranspiration calculation method according to the type of hydrological mechanism model used.

[0140] Furthermore, long-term water balance constraints are used to intercept spurious water consumption or unauthorized increases in the water volume of the interception mechanism model. In actual calculations of hydrological processes, the accurate selection of the evaporation variable has a decisive impact on the system's water volume closure. If the potential evapotranspiration is substituted into the balance equation for calculation, since the potential evapotranspiration represents the maximum evaporation capacity under unlimited water supply conditions, its value is usually much larger than the actual evaporation loss occurring at the surface. This operation will cause the calculated system closure residual to deviate from the true physical laws, resulting in a large number of erroneous deletions of physically feasible sequences. Therefore, this implementation method forcibly obtains and substitutes the actual evapotranspiration, which reflects the true water supply constraints of the underlying surface, for deduction, in order to limit the true water cycle closure boundary. The long-term water balance calculation logic is as follows:

[0141] ε wb =|ΣP eff,t -ΣET a,t -ΣQ xaj-snow,t -ΔW| / ΣP eff,t ;

[0142] Where, ε wb For the calculated relative ratio, ΣP eff,t ΣET represents the cumulative total amount of effective liquid water over a multi-year timescale. a,t ΣQ represents the cumulative total of the actual evapotranspiration within the corresponding time period, calculated hour by hour. xaj-snow,t ΔW represents the cumulative total of simulated total runoff over the corresponding time period, ΔW represents the change in watershed water storage calculated by the difference in soil moisture content state variables at both ends of the corresponding evaluation period, and Σ() represents the operation of summing up continuous time series variables over a multi-year time scale.

[0143] For example, assuming that during the selected multi-year evaluation period, the system cumulatively receives a total effective liquid water volume of 1000.0, acquires and accumulates an actual evapotranspiration of 400.0, extracts a cumulative simulated total runoff of 550.0, and calculates the watershed water storage change from the initial and final soil moisture content as 20.0. Substituting these values ​​into the above calculation logic, the system closure residual is calculated as: 1000.0 - 400.0 - 550.0 - 20.0 = 30.0. The relative ratio between the closure residual and the total effective liquid water volume is calculated as: 30.0 / 1000.0 = 0.03. Under this calculation result, if the preset balance threshold is set to 0.05, since the ratio 0.03 < 0.05, the system determines that the characteristic trajectory generated by this set of parameters meets the water balance closure requirement and retains it. The above preset balance threshold can be adaptively adjusted by those skilled in the art based on the data quality and observation accuracy of the target watershed; those skilled in the art can determine a suitable value without creative effort.

[0144] In one possible implementation, the verification logic for screening using source runoff conservation constraints specifically includes:

[0145] The summation of surface runoff, soil runoff, and groundwater runoff at the current moment yields the sum of the corresponding fast and slow flow components;

[0146] Calculate the absolute residual between the sum of fast and slow flow components and the simulated total runoff, and determine that the corresponding characteristic trajectory passes the source runoff conservation constraint when the absolute residual is less than the preset deviation tolerance threshold.

[0147] Specifically, the source runoff conservation constraint is used to control the consistency of the multi-source addition logic in the runoff calculation process of the control mechanism model. Due to the strong nonlinear linkage effect of parameters within the watershed in high-altitude cold regions, some randomly combined runoff generation and confluence parameters may cause numerical oscillations and loss of individual runoff components during mathematical iteration calculations. This results in a residual that deviates from physical principles between the sum of surface and groundwater components and the uniformly calculated total runoff. This verification calculation process uses the following absolute deviation relationship for judgment and interception:

[0148] res abs =|R s,t +R i,t +R g,t -Q xaj-snow,t |;

[0149] Among them, res abs R is the calculated absolute residual. s,t To characterize the surface runoff at the current moment, R i,t R is the soil flow rate at the current moment. g,t Q is the current time-specific groundwater runoff. xaj-snow,tThe simulated total runoff at the current moment is used to characterize the total runoff.

[0150] In specific computational scenarios, the system extracts and merges independent source flow rates on a time-by-time basis. Assuming that the surface runoff extracted in a certain time slice is 1.2, the soil runoff is 0.5, and the groundwater runoff is 0.3, the summation of the fast and slow flow components yields a total of 2.0. If the simulated total runoff record extracted in the same time slice is 1.995, the total difference between the two characteristic states is calculated, and the absolute value is taken to obtain an absolute residual of 0.005. If the preset deviation tolerance threshold is set to 0.01, since the calculated absolute residual of 0.005 is strictly less than the 0.01 threshold, the system determines that the source water quantity relationship at this time point conforms to the principle of mass addition, and allows the data features at this moment to be stored in the valid sequence pool.

[0151] In some alternative implementations, for cascaded watershed networks with large lake storage capacity or reservoir projects, after performing the above-mentioned summation steps for fast and slow flow components, a predetermined lake node storage difference variable can be added to compensate for the delayed storage deviation caused by large-scale still water areas, so that the calculation boundary of the absolute residual can be compatible with more complex surface hydrological topological conditions.

[0152] Example 5: Based on the above examples, this example details the process of inputting ensemble features into a pre-trained time-series mapping network to output simulated runoff results. Specifically, it elucidates the offline statistical construction process addressing uncertainties in multiple trajectory parameters, and the multi-evaluation metric conflict resolution mechanism at the end of deep learning network training.

[0153] In one possible implementation, such as Figure 4 As shown, before the step of inputting the set features into the pre-trained temporal mapping network, there is also a pre-training process for offline construction of the deep learning model:

[0154] Extract the values ​​of multiple candidate feature trajectories on the same physical state variable at each time step and calculate the mean;

[0155] Calculate the sum of squares of the differences between the corresponding variable values ​​and the mean of each candidate feature trajectory, and extract the arithmetic square root of the sum of squares after dividing by the total number of trajectories to obtain the standard deviation characterizing the dispersion of the multi-trajectory distribution;

[0156] The mean and standard deviation are concatenated and fused to form a set feature sequence, and the set feature sequence is truncated according to a set sliding time window scale to construct supervised training samples; further, it can be implemented as follows: the mean and standard deviation are concatenated and fused to form a set feature sequence, and the set feature sequence is truncated according to a set sliding time window scale, while the pre-acquired historical measured runoff data at the corresponding time is used as training labels and paired with the set feature sequence to construct supervised training samples;

[0157] The weight parameters of the initial deep learning network are iteratively updated using supervised training samples, thereby obtaining a pre-trained temporal mapping network.

[0158] Specifically, the multiple candidate feature trajectories retained after consistency filtering constitute a set of physically feasible parameter solution spaces. Conventional fusion calculations often discard information from multiple solutions, selecting only a single optimal parameter trajectory for training, causing the model to lose its ability to perceive the inherent uncertainty of hydrophysical evolution. To transform this uncertain parameter into an explicit supervised input, this step implements dimensionality reduction statistical feature encoding. When extracting the time series of predetermined physical state variables, the mean distribution and discrete boundaries across trajectories are obtained using the following linear combination rule:

[0159] ;

[0160] in, Let X be the mean of the predetermined state variables at time t. t k Let Σ be the value of the corresponding variable for the k-th candidate feature trajectory at time t, K be the total number of candidate feature trajectories to be selected and retained, and Σ() be the summation operation performed on the values ​​of all K trajectories at the same time.

[0161] After the mean convergence is achieved, the system synchronously calculates the dispersion index for each time step. To ensure that the extracted dispersion feature scale maintains strict dimensional consistency with the original physical state variables, an arithmetic square root operation must be applied to the calculated statistical variance:

[0162] ;

[0163] Where, σ X,t Let the standard deviation of the predetermined state variable be given at time t. Let Σ be the square of the difference between the value of the k-th candidate feature trajectory and the mean, Σ() be the cumulative sum of the squared differences of all trajectories, and K be the total number of trajectories.

[0164] The mean and standard deviation derived from each physical state variable are orthogonally concatenated along the data channel dimension. For a single physical moment, the original eleven-dimensional physical state variables are expanded and integrated into a twenty-two-dimensional combined channel after the above statistical concatenation. Further, a data truncation window is initiated to perform synchronous truncation along the time axis. Taking a sliding time window of 30 time steps as an example, 30 consecutive state records, including the current node, are truncated, thereby reconstructing the one-dimensional set of features into a two-dimensional input matrix with a dimension scale of (30, 22). This matrix carries the evolution trend of hydrological features within the window period and the corresponding uncertainty interval amplitude, constituting a complete supervised training sample.

[0165] After acquiring supervised training samples, the system imports them into a Long Short-Term Memory (LSTM) network to process time series data. The initial deep learning network employs a temporal topology with forgetting and update gating mechanisms. In the offline pre-training phase, the network is configured with an architecture containing two hidden layers, each with a neuron dimension of 64, and a dropout rate parameter of 0.2 to prevent overfitting. During iterative optimization, the loss deviation between the predicted sequence and the measured runoff is calculated, and the error gradient is backpropagated to the network nodes using the Adam (Adaptive Moment Estimation) optimization algorithm. For example, the initial learning rate is configured to 0.005, and the batch size for a single gradient update is 0.1% of the total training pool. Through a fixed number of iterations, the network node weights gradually fit the nonlinear manifold mapping relationship of the physical multi-track uncertainty to the final runoff, thereby outputting a fixed pre-trained temporal mapping network.

[0166] Furthermore, during the iterative update of the weight parameters of the initial deep learning network using supervised training samples, each neuron node of the initial deep learning network is nonlinearly gated using hyperbolic tangent activation function and sigmoid (S-shaped) activation function. At the same time, the system specifies the mean square error loss function as the quantitative evaluation standard for loss bias. By calculating the mean square error between the predicted runoff sequence corresponding to the twenty-two-dimensional combined channel input in the supervised training samples and the historical measured runoff labels, the error gradient is backpropagated to the network nodes through the Adam adaptive optimization algorithm until the network converges or reaches the maximum number of training rounds, thereby obtaining the pre-trained temporal mapping network.

[0167] It should be noted that the above-described method of constructing a 22-dimensional combined channel input feature through statistical splicing explicitly represents the intrinsic physical correlation between the input data and the final output simulated runoff. Specifically, the mean feature extracted time-by-time represents the central trend of the physical evolution mechanism of the target watershed under the interference of hydrological model parameter uncertainties, while the simultaneously extracted standard deviation feature qualitatively represents the boundary amplitude of physical parameters caused by plateau microclimate disturbances. The Long Short-Term Memory (LSTM) network, through a temporal gating mechanism, simultaneously inputs and learns the temporal combination of mean and standard deviation, enabling it to establish a nonlinear manifold mapping relationship in its internal neuron weights of the propagation of uncertainty from source water input to outlet cross-section runoff. This demonstrates a rigorous mechanistic intrinsic correlation between the fused feature input system and the final simulated runoff results, providing sufficient physical interpretability support for the online adaptive generalization extrapolation of the entire data stream.

[0168] One possible implementation includes, after obtaining the pre-trained temporal mapping network, a step of selecting the final model using a multi-metric trade-off mechanism:

[0169] Using historical observation data from the pre-acquired independent validation period, the Nash efficiency coefficient and auxiliary evaluation index of the pre-trained temporal mapping network under different training rounds are calculated. The auxiliary evaluation index includes at least the root mean square error and the mean absolute error.

[0170] When the Nash efficiency coefficient and the auxiliary evaluation index change in the same trend, the network state corresponding to the training round with the optimal Nash efficiency coefficient is taken as the preferred result.

[0171] When the Nash efficiency coefficient deviates from the trend of auxiliary evaluation indicators, the evaluation conflict handling mechanism is triggered.

[0172] Under the evaluation conflict handling mechanism, the runoff simulation error of the corresponding network state during the snowmelt sensitive period is extracted, and the network state with lower runoff simulation error is used as the priority discrimination criterion to complete the final selection of the pre-trained time-series mapping network.

[0173] Furthermore, a multi-index trade-off mechanism is used to address the evaluation mismatch phenomenon commonly seen in complex nonlinear networks during long-term time-series regression tasks. At the end of the parameter optimization phase of the pre-trained time-series mapping network, a retained independent validation period sequence is used for performance verification. Multiple conventional evaluation functions are computed in parallel at each detection step. Since each evaluation function has a different sensitivity to the error sequence—for example, the Nash efficiency coefficient is more biased towards fitting the characteristics of high-water peaks, while the root mean square error and mean absolute error tend to measure the mean drift over the entire time period—the extension of network training may make it difficult to achieve simultaneous global optimization for these multiple indices. The training trend is monitored by continuously comparing the differences in various indices generated in adjacent rounds. When an increase in the Nash efficiency coefficient is detected, while the variability ratio or root mean square error used to verify distribution consistency shows signs of deterioration and divergence, it is determined that the evaluation index trend has deviated.

[0174] After triggering the evaluation conflict handling mechanism, the system abandons the conventional pure mean truncation strategy and initiates a local time-priority verification logic tailored to the characteristics of high-altitude and cold regions. It locates and extracts the snowmelt-sensitive period interval from the hydrological calendar, specifically covering the spring to early summer uplift period within the natural year. The daily runoff simulation deviation within this specific period is calculated. The comprehensive absolute error of the snowmelt-sensitive period is ranked above all indicators across the entire time period, and the network weight snapshot with the smallest margin of deviation within this interval is selected and defined as the final model parameter set. This avoids the pitfall of the model excessively sacrificing the prediction accuracy of the local snowmelt uplift period to cater to the overall annual evaluation indicators.

[0175] In some alternative implementations, for remote, high-altitude sub-basins lacking long-sequence measured runoff records, the priority criteria in the evaluation conflict resolution mechanism can be replaced by feature substitution. This involves introducing the inversion time series of external elevation snow cover remote sensing products as an alternative verification target, and using the snow retreat rate error derived from remote sensing as a reference variable for adjudication when evaluating conflicts.

[0176] Example 6: Based on the same inventive concept as the above examples, this example provides another method for obtaining the candidate parameter set of a hydrological mechanism model. The specific difference lies in applying a priori physical boundary constraints before performing consistency screening. The remaining steps are the same as in the previous examples and can be referred to the foregoing description; they will not be repeated here.

[0177] In one possible implementation, prior to the step of filtering the sequence of the multidimensional fusion feature system based on a preset physical consistency condition, the method further includes:

[0178] A random sampling strategy is used to generate multiple sets of candidate parameters for hydrological mechanism models, and physical boundary rules are set during the sampling and generation process. Each set of candidate parameters is used to configure and drive the hydrological mechanism model, resulting in a sequence of corresponding multi-dimensional fusion feature systems.

[0179] The physical boundary rules include: limiting the sum of the runoff generation coefficient representing subsurface runoff and the runoff generation coefficient representing interflow to be strictly less than a preset combined boundary threshold, in order to eliminate unreasonable combinations of runoff generation parameters from different sources.

[0180] Specifically, the process of generating multiple sets of candidate parameters for the hydrological mechanism model using a random sampling strategy is used to provide the model with a divergent parameter solution space. The hydrological mechanism model contains numerous empirical parameters controlling the evolution of runoff generation and confluence. Directly substituting these parameters into an unconstrained random number set has a high probability of causing computational overflow or physical evolution paradoxes. The random sampling strategy is implemented by performing uniformly distributed sampling within the physically reasonable range of each hydrological parameter. Since groundwater runoff and interflow both belong to the groundwater system's drainage process below the surface, they share the infiltration water quota of the underlying surface. Therefore, there is a physical water quantity competition between the runoff generation coefficient representing groundwater runoff and the runoff generation coefficient representing interflow. If both obtain high values ​​during random generation, the total water volume allocated to the underground space will exceed the actual maximum infiltration capacity at the predetermined soil depth, leading to a calculation error where the total runoff coefficient exceeds the boundary. To prevent such invalid parameter sets from entering the subsequent time-series matrix operation stage, the system synchronously triggers the conditional judgment verification of the physical boundary rules during the sampling generation cycle.

[0181] The calculation relationship between the sum of the runoff generation coefficient representing subsurface runoff and the runoff generation coefficient representing interflow in the above-mentioned limited actions can be expressed as follows:

[0182] K sum =Kg+Ki;

[0183] Among them, K sum The sum of the calculated runoff coefficients is given by Kg, which is the runoff coefficient characterizing groundwater runoff (groundwater runoff outflow coefficient), and Ki, which is the runoff coefficient characterizing interflow (interflow outflow coefficient).

[0184] Furthermore, the sum of the calculated flow generation coefficients is compared with a preset combined boundary threshold. The corresponding judgment relationship is as follows:

[0185] K sum <T boundary ;

[0186] Among them, K sum T is the sum of the calculated flow generation coefficients. boundary This is a preset combined boundary threshold.

[0187] When the above judgment relationship is true, the system retains the currently generated set of candidate parameters for the hydrological mechanism model and stores it in the subsequent operation call queue; when the above judgment relationship is false, the system performs a removal action to destroy the set of parameters and immediately triggers the next round of resampling.

[0188] As an example, the preset combined boundary threshold can be configured to 0.8, taking into account the permeability characteristics of the permafrost layer in high-altitude cold regions. Assume that the system obtains a runoff coefficient of 0.45 representing subsurface runoff and a runoff coefficient of 0.40 representing interflow runoff during a single sampling process. Substituting these values ​​into the addition relationship, the sum of the runoff coefficients is calculated to be 0.85. A comparison and judgment operation is then performed. Since the value 0.85 is greater than the boundary value 0.8, it does not satisfy the strict less-than judgment relationship, and this parameter combination is identified as an unreasonable source runoff parameter combination and is discarded by the system. Assume that after triggering resampling, the system obtains a runoff coefficient of 0.30 representing subsurface runoff and a runoff coefficient of 0.25 representing interflow runoff. Substituting these values ​​again, the sum of the runoff coefficients is calculated to be 0.55. Since the value 0.55 is strictly less than the boundary value 0.8, it satisfies the judgment relationship, and the system retains this set of candidate parameters.

[0189] In some alternative implementations, in addition to limiting the sum of the two runoff generation parameters mentioned above, the physical boundary rules can also be configured with absolute extreme value constraints for individual hydrological characteristic parameters. Constraint rules can be added, such as limiting the conversion factor characterizing deep evaporation in the watershed to fluctuate within the range of 0.1 to 1.5, or limiting the vadose zone tension water capacity to be greater than a predetermined numerical baseline. Adding such parallel absolute numerical boundary constraints can further compress the meaningless random sampling search space and reduce the computational resource load when performing subsequent multidimensional quantitative physical consistency screening steps.

[0190] Example 7: Based on the same inventive concept as the above examples, another pre-implementation method for data acquisition and processing is provided. The specific difference lies in the introduction of unified data cleaning and variable transformation rules after acquiring the driving data. The remaining steps are the same as in the previous examples and can be referred to the foregoing description; they will not be repeated here.

[0191] In one possible implementation, the step of acquiring meteorological driving data for the target watershed also includes a unified data preprocessing procedure:

[0192] Missing records in the meteorological driving data are uniformly filled with zero-value boundaries, and extreme value parameters for all time periods are used to perform variable-wise normalization mapping on each variable in the meteorological driving data.

[0193] Furthermore, the extreme value parameters in the normalized mapping are strictly kept consistent throughout the entire time period during both the offline pre-training stage and the real-time simulation output stage.

[0194] Specifically, the acquisition of meteorological driving data often relies on long-term continuous observations from field stations. Data gaps are inevitable due to instrument malfunctions or communication interruptions. In practical engineering applications in high-altitude and cold-region watersheds, missing data are typically marked as invalid values. For these invalid values, the system performs a mandatory zero-value boundary filling operation. This processing logic takes into account that, for precipitation, a zero value represents no precipitation event, preventing the introduction of false water volumes; for temperature and evaporation, resetting them to a baseline zero point prevents gradient explosion during backpropagation in deep neural networks due to extreme outlier inputs, ensuring the convergence stability of the time-series mapping network.

[0195] After completing the missing value imputation, extreme value normalization mapping is then performed on the meteorological driving data and all subsequent derived state variables. This compresses data with different physical dimensions and numerical distribution ranges into a dimensionless unified scale, eliminating the dominant interference of high absolute value variables on model weight updates.

[0196] The linear calculation formula for the normalization action is as follows:

[0197] X norm,t=(X t -X min ) / (X max -X min );

[0198] Among them, X norm,t Let X be the dimensionless eigenvalue at time t after normalization mapping. t Let X be the original physical variable measurement value at time t. min Let X be the parameter representing the minimum value of this predetermined variable across all historical sample sets over all time periods. max This is the maximum value parameter of the predetermined variable across all historical sample sets.

[0199] It should be noted that the extreme value parameters obtained by the system for all time periods must be extracted from the entire sequence of the training set, and the same set of extreme value parameters must be forcibly reused when performing feature transformations for independent validation periods or real-time forecast periods.

[0200] Suppose that the highest recorded temperature in a high-altitude basin during a pre-defined training period is 25.0°C and the lowest is -35.0°C. If the temperature on a given day is 10.0°C during the real-time prediction phase, the normalized value calculated using the extreme value parameters is: (10.0 - (-35.0)) / (25.0 - (-35.0)) = 0.75. By strictly maintaining parameter consistency, the system cuts off the feedback path of future data to model training, ensuring that the physical causal logic learned by the mapping network is not interfered with by false information.

[0201] As an alternative to the aforementioned deep learning network topology, if in step 103, the physical state variables are not extracted using a hydrological mechanism model, and consistent trajectory screening is not performed, but instead the pure meteorological sequence after the above extreme value normalization process is directly input into the network, a purely data-driven baseline network model is formed. This baseline model can usually quickly reduce absolute error on the training set, but in the validation phase, due to the lack of physical constraints on the delayed response of snowmelt and internal reservoir capacity, its output runoff curve will show significant phase lag and flood peak underestimation during the snowmelt sensitive period. By comparing the error residual sequence of this alternative baseline scheme with the multi-level feature fusion scheme disclosed in this invention, the specific gain effectiveness of introducing physical mechanisms can be intuitively verified.

[0202] Example 8: Based on the same inventive concept as the above examples, this example provides an extended implementation method for the operational application of runoff simulation methods for watersheds in high-altitude and cold regions. The above examples focused on the data processing and computational extrapolation process of the runoff simulation model. This example further explains how to apply the runoff prediction data output by the above examples to actual engineering scheduling and disaster prevention and mitigation scenarios. The remaining network training and extrapolation steps are the same as in the previous examples and can be referred to the foregoing description; they will not be repeated here.

[0203] In one possible implementation, after the step of outputting the simulated runoff results, an operational application process is also included:

[0204] The simulated runoff results are input into the watershed water resources scheduling system or flood warning platform; triggering the watershed water resources scheduling system or flood warning platform to generate hydrological scheduling strategies or flood hazard warning instructions for the snowmelt sensitive period in high-altitude and cold regions.

[0205] Specifically, the simulated runoff results are represented in computer memory as a time-series data array containing daily or hourly flow forecasts for a predetermined future time window. High-altitude watersheds are sensitive to snowmelt during spring and early summer, when rising temperatures accelerate the melting of snow and deep glaciers, easily triggering sudden snowmelt floods or providing concentrated agricultural irrigation water for downstream areas. The simulated runoff results are then fed into a watershed water resources management system or flood warning platform to transform the virtual data array generated by the algorithm model into executable engineering control signals or disaster response actions in the physical world.

[0206] Furthermore, the watershed water resources scheduling system and flood early warning platform can be a computer cluster or cloud server deployed at the watershed management agency, which integrates reservoir area curves, gate discharge constraint functions, and river evolution hydrodynamic models. After acquiring the above input data, the system performs boundary condition comparisons and evolution calculations.

[0207] In water resource allocation scenarios, the process of generating hydrological allocation strategies specifically includes: obtaining the current water level and available flood control capacity of key reservoirs within the basin, and inputting simulated runoff results as boundary inflow conditions into the reservoir's water balance equation. If the calculated total runoff within a predetermined period will exceed the current reservoir's flood control storage limit, a hydrological allocation strategy to increase the discharge flow is triggered. If the calculated total runoff is within a safe storage range and there is a downstream irrigation water shortage, a hydrological allocation strategy to store water or release water stably as needed is triggered. For example, when the simulated runoff results for the next 15 days show a clear monotonic upward trend, the water resource allocation system calculates and determines that the current reservoir capacity is insufficient to handle the expected snowmelt flood peak, and then outputs physical control command parameters for early gate opening and pre-discharge. These command parameters specifically include the predetermined discharge flow value, the number of gates to be opened, and the reserved duration.

[0208] In flood warning scenarios, the process of generating flood hazard warning instructions specifically includes: comparing the peak flow characteristics in the simulated runoff results with the corresponding flow thresholds for historical warning water levels at downstream key river sections. When the simulated peak flow value is determined to be greater than the preset warning flow threshold, the system triggers the flood control response mechanism and generates a flood hazard warning instruction. The flood hazard warning instruction includes the estimated time for the flood to reach the predetermined spatial section, the estimated magnitude of the excess flow, and the corresponding risk level indicator data.

[0209] In some alternative implementations, the aforementioned docking input process can achieve real-time data transmission and parsing through the construction of a standardized application programming interface (API). Furthermore, as an alternative or extended solution, the simulated runoff results can also be synchronously docked into the hydropower generation dispatching and control system. By combining the upstream and downstream head difference data of the cascade reservoirs with the generator unit output characteristic curves, an active power load allocation and dispatching strategy for the turbine units can be generated to maximize the power generation benefits across the entire basin.

[0210] The implementation process of this invention in diurnal runoff simulation is illustrated using a specific watershed as an example. The watershed area is approximately 140,000 km². 2 The runoff supply is controlled by both rainfall and snowmelt, exhibiting a distinct seasonal snowmelt supply characteristic. The data used in this embodiment include rainfall, potential evapotranspiration, average temperature, and observed runoff, spanning from 1980 to 2024, with a daily time resolution. The average temperature is preferably the basin-wide average daily temperature obtained after area-weighted or zonal processing of the stations; for basins with large elevation differences, this simplification is applicable to engineering applications under daily-scale and area-weighted conditions.

[0211] According to one aspect of this application, the technical method and implementation process of the present invention are described by following the workflow of data preprocessing - snowmelt process construction - fusion feature system construction - candidate trajectory screening - deep learning fusion training - result output. A deep learning runoff simulation method for rain-snow-melt mixed recharge watersheds is also described. Figure 6 As shown, it may also include the following steps:

[0212] Step S1 involves collecting, quality controlling, and preprocessing sample data.

[0213] Step S1.1 involves data collection and compilation. Collect continuous daily-scale rainfall, potential evapotranspiration, average temperature, and observed runoff data from the outlet control station for the target watershed. Unify data from different sources to the same time reference and form a continuous drive-response sample set based on daily sequences. Preferably, rainfall, evapotranspiration, and temperature are measured values ​​from stations or spatially weighted watershed averages, and runoff is measured flow sequences from the outlet section.

[0214] Step S1.2: Perform consistency checks and missing data handling. The original time series is checked for completeness and consistency. For records marked with missing data as -9999.9, -9999, -99, or -99.9, in this embodiment, they are uniformly replaced with 0.0, and the remaining empty values ​​are filled with 0.0. Subsequently, each variable is normalized using a variable-wise min-max normalization (maximum-minimum normalization). The normalization parameters are the minimum and range of the corresponding variables across all time periods, and are consistent between the training and prediction phases.

[0215] Step S1.3 involves sample partitioning and time series construction. The training and validation periods are divided in a 7:3 ratio according to time sequence to ensure that validation period samples do not participate in model parameter updates. In this embodiment, the input sequence length L=30d, and a sliding window method is used to construct supervised learning samples. The sample at time t consists of multi-feature sequences within the time period [t-L+1,t], and the corresponding output is the target runoff at time t.

[0216] Furthermore, in this embodiment, the division between the training period and the validation period can be achieved by continuously dividing the time sequence. When using long-sequence daily-scale samples from 1980 to 2024, the training period samples are located in the first part of the time series, and the validation period samples are located in the second part, thereby avoiding information leakage.

[0217] Step S2: Construct a snowmelt module (snowmelt dynamic mechanism model) that includes rain and snow segmentation, snow accumulation, snow-constrained snowmelt, refreezing, and snow water equivalent update, and couple it with the Xin'anjiang model.

[0218] Step S2.1: Establish the rain-snow separation relationship. Let the total precipitation at time t be P. t The average temperature is T t The pure snow temperature threshold is T s The pure rain temperature threshold is T r .

[0219] when At that time, P rain,t =0, P snow,t =P t ;

[0220] when At that time, P rain,t =P t P snow,t =0;

[0221] When T s <T t <T r hour, , .

[0222] Preferably, a snowfall correction factor (SFCF) can be introduced to correct the measurement bias of solid precipitation; in this embodiment, T s T r The SFCF is determined by parameter calibration.

[0223] Step S2.2: Establish the relationship between snow accumulation-constrained snow melting and refreezing.

[0224] First calculate the water equivalent of intermediate snow accumulation. ;

[0225] Calculate potential snowmelt and according to the actual amount of snow melted Implement snow restraints;

[0226] Simultaneously calculate the potential refreezing volume. And according to the actual amount of refreezing Liquid water confinement is implemented.

[0227] These constraints prevent the generation of false snow melt volume under snowless conditions and limit the refreezing volume to the upper limit of available liquid water.

[0228] Step S2.3: Update the snowmelt equivalent and construct effective liquid water. The snowmelt equivalent is calculated according to... Dynamically updated; effective liquid water entering the Xin'anjiang model according to... Calculate and satisfy the water conservation relationship. .

[0229] This enables a continuous description of the processes involved in temperature-controlled precipitation phases, snow storage, snowmelt release, and liquid water infiltration and runoff generation.

[0230] In this embodiment, the parameters of the snow melting module and the parameters of the Xin'anjiang model are both generated by uniformly distributed random sampling to form candidate parameter groups. The lower limit constraint of Ki is a positive value, and Ki and Kg together satisfy the physical constraint condition Kg+Ki<0.8 to avoid unreasonable source combinations.

[0231] Step S2.4: Couple the snowmelt module with the three-source Xin'anjiang model to obtain the XAJ-SNOW physical model. Using the effective liquid water obtained in step S2.3 as the runoff input for the Xin'anjiang model, sequentially complete the evapotranspiration calculation, full-storage runoff calculation, three-source allocation, and confluence calculation to obtain simulated runoff and physically meaningful internal state variables. Preferably, the Xin'anjiang model parameters satisfy the physical feasibility constraints.

[0232] Examples of parameter values ​​in this embodiment are as follows:

[0233] The evapotranspiration conversion factor Kc is taken as 0.8, the surface free water storage capacity SM is taken as 35mm, the groundwater outflow coefficient Kg is taken as 0.3, the soil outflow coefficient Ki is taken as 0.2, the groundwater receding coefficient Cg is taken as 0.98, the soil receding coefficient Ci is taken as 0.8, the river network water storage receding coefficient CR is taken as 0.35, the upper tension water storage capacity WUM is taken as 100mm, the lower tension water storage capacity WLM is taken as 50mm, the deep tension water storage capacity WDM is taken as 200mm, the impermeable area ratio Imp is taken as 0.05, the water storage capacity curve index B is taken as 0.9, the deep evapotranspiration coefficient C is taken as 0.12, the free water storage capacity curve index Ex is taken as 1.2, the river network lag time LAG is taken as 2, and the slope confluence coefficient CS is taken as 0.35, and Kg+Ki<0.8 is satisfied.

[0234] The above parameter values ​​can be determined by those skilled in the art based on comprehensive calculations using watershed hydrological and meteorological data and model parameter calibration methods. This invention is not limited to the specific values ​​mentioned above.

[0235] Step S3: Construct a fusion feature system that covers the entire physical chain.

[0236] Step S3.1: Run the coupled XAJ-SNOW physical model to extract the simulated total runoff. Soil moisture content Surface runoff , soil in the stream underground runoff Effective liquid water , snowmelt amount and snow reserves Process variables, etc. These process variables are used to characterize information on runoff generation, source distribution, and snowmelt replenishment, and serve as physical constraint feature inputs for subsequent deep learning fusion training.

[0237] Step S3.2, classify the original driving factors and physical process variables according to physical level, specifically including: external driving layer [ , , Snow melting transition layer , , ], water storage state layer [ Source-specific response layer , , , In this invention, the hierarchical physical process feature chain refers to a feature-based system covering the above four physical levels, used to characterize the entire process of rain and snow-snowmelt-water storage-source response in high-altitude and cold regions.

[0238] Step S3.3, this embodiment uses eleven-dimensional input features:

[0239] X t =[ , , , , , , , , , , ].

[0240] In addition to meteorological driving variables, the eleven-dimensional input features also include effective liquid water, snowmelt, snow storage, soil moisture content, and source response variables, in order to form a fusion input feature system for mixed recharge basins in high-altitude and cold regions.

[0241] The time subscript t represents the variable at time t; where P t For total precipitation, For potential evaporation, T t The average temperature For effective liquid water, This represents the actual amount of snow melted. Equivalent to snowmelt Soil moisture content, To simulate total runoff, , and These refer to surface runoff, interflow, and groundwater runoff, respectively. Where there is no ambiguity, these symbols can also represent the corresponding variable sequences.

[0242] Step S3.4: To enhance sample usability, the above variables are simultaneously standardized, and the same transformation parameters are used in both the training and prediction phases. For variables with significant differences in magnitude, a variable-wise min-max normalization method is preferred to avoid high-magnitude variables dominating network weight updates.

[0243] Step S4: Candidate physical trajectory generation and physical consistency constraint screening.

[0244] Step S4.1: To reduce the impact of physical model parameter uncertainty on the fusion model, multiple candidate parameter sets are generated using a Monte Carlo random sampling strategy within the physically reasonable range of each parameter.

[0245] In this embodiment, the pure snow temperature threshold Ts is set to -1.0℃, the pure rain temperature threshold Tr is set to 2.0℃, the snowmelt initiation temperature T0 is set to 0.5℃, the degree-day factor DDF is set to 4.0 mm / (℃·d), and the refreezing coefficient C is set to... rfThe snowfall correction factor (SFCF) is set to 0.05, and the SFCF is set to 1.1. These parameters can be determined by those skilled in the art based on historical meteorological and hydrological observation data of the target watershed, using Monte Carlo random sampling combined with physical consistency constraints.

[0246] In the embodiments, the number of sampling groups can be set according to the computing power conditions, for example, 1000 groups, and 50 to 200 candidate trajectories are retained after screening; if a code-level rapid experiment is used, a smaller number of groups can also be used for preliminary verification.

[0247] Step S4.2: Drive the XAJ-SNOW physical model using each set of candidate parameters to obtain the corresponding simulated total runoff and internal process variable sequences, and perform physical consistency screening based on the following quantitative criteria:

[0248] ① Long-term water balance constraints, calculation Preferred requirements ≤0.05;

[0249] ② Variable non-negativity constraint, requiring , , , , , Non-negative throughout the entire time period;

[0250] ③ Seasonal snowmelt phase constraints require that the peak snowmelt month in the cold regions of the Northern Hemisphere fall between March and July;

[0251] ④ Source runoff conservation constraints, requiring The preferred value is δ=0.01mm / d;

[0252] ⑤ Minimum performance threshold: The XAJ-SNOW physics model must achieve an NSE of at least 0.30 during training. Only candidate trajectories that simultaneously meet the above criteria are retained.

[0253] Step S4.3: Perform time-by-time statistical aggregation on the K candidate trajectories and their corresponding feature sequences retained in step S4.2; for any process variable... Calculate the mean at each time point. and standard deviation This is to form a set of features that simultaneously characterizes central tendency and dispersion.

[0254] It should be noted that S4 is only responsible for the preliminary screening of the physical consistency of candidate trajectories and the construction of set features. It does not undertake the deep learning model training and final selection functions to avoid process overlap with subsequent steps.

[0255] Step S5: Conduct hydrological model-deep learning fusion training based on multi-trajectory set feature encoding.

[0256] Step S5.1: Establish an LSTM (Long Short-Term Memory) network structure. The ensemble feature sequence obtained in step S4.3 is used as the network input, and the target runoff is used as the output. The ensemble features include at least the time-by-time mean features of each physical process variable, and preferably further include the corresponding standard deviation features, to simultaneously characterize the central tendency and dispersion of the set of physically feasible trajectories. In this embodiment, the network has 2 hidden layers, 64 hidden units, and a dropout rate of 0.2. The aforementioned hyperparameters, such as the number of hidden layers, the number of neurons, and the dropout rate, can be determined by those skilled in the art through conventional hyperparameter tuning methods such as cross-validation or grid search; this invention is not limited to the specific values ​​mentioned above.

[0257] Step S5.2: Formulate the training strategy. The mean squared error (MSE) is used as the loss function, the Adam optimizer is used for parameter updates, and a learning rate decay strategy is applied to improve convergence stability. In this embodiment, the initial learning rate is 0.005, the maximum number of training epochs is 1000, the batch size is 1 / 10 of the number of training samples, and the learning rate decay parameters are a step size of 100 and a decay factor of 0.5. Starting from the 100th epoch, the validation period Nash efficiency coefficient (NSE) is monitored, and the model state corresponding to the maximum NSE during the validation period is saved. Since the network input already contains statistical aggregation information of multiple physically feasible trajectories, it is not necessary to train an independent LSTM for each candidate trajectory separately, which can reduce training overhead and information waste while maintaining physical constraints.

[0258] Step S5.3: Select evaluation indicators. The Nash efficiency coefficient (NSE), Kling-Gupta efficiency coefficient (KGE), root mean square error (RMSE), and mean absolute error (MAE) are used to evaluate model performance. NSE is the primary preferred indicator, KGE is used to comprehensively assess correlation, variability, and consistency of bias, and RMSE and MAE reflect the magnitude of the error. For cases where the improvement directions of KGE and NSE are inconsistent, their correlation component (r), variability ratio (α), and bias ratio (β) should be further analyzed to explain the trade-offs between different evaluation indicators.

[0259] Step S6: Model optimization and runoff output.

[0260] Step S6.1: The candidate fusion models trained in step S5 are then optimized based on the principle of prioritizing NSE during the validation period, supplemented by KGE, RMSE, and MAE, to output the diurnal runoff process for the target time period. Since the input features explicitly include P... eff,t M t SWE t The model incorporates source variables from the Xin'anjiang model and further integrates statistical aggregation information from multiple physically feasible trajectories. Therefore, the model can more fully perceive the impact of snow and ice replenishment in high-altitude and cold regions on runoff formation and its parameter uncertainties.

[0261] Step S6.2, as follows Figure 7 As shown, the implementation results indicate that, compared with the traditional LSTM that only uses the original meteorological drive, the XAJ-LSTM (Long Short-Term Memory Network-Xin'anjiang Model) has shown some improvement in the validation period after introducing the Xin'anjiang physical process variables; after further explicitly introducing rain and snow segmentation, snow accumulation and snowmelt release information, the XAJ-SNOW-LSTM (Long Short-Term Memory Network-Xin'anjiang Coupled Snowmelt Module Model) has the best overall performance.

[0262] Table 1 Evaluation of indicators during model validation period

[0263]

[0264] The ablation comparison results shown in Table 1 indicate that although the KGE of both XAJ-LSTM and XAJ-SNOW-LSTM are lower than that of the baseline LSTM, the method of this invention still has better implementation effects in terms of NSE, RMSE, and MAE indices, particularly in terms of peak flow fitting, identification of snowmelt-sensitive flow rise process, and overall error control. If the KGE is lower than the baseline, it can be explained by the sequence variability ratio deviating from 1 while the model improves the fitting of peak flow and snowmelt-sensitive process. This phenomenon reflects the trade-off between different evaluation indices and does not negate the technical effectiveness of this invention in the target scenario. Based on the above, the preferred implementation method is multi-trajectory set feature encoding, which can further utilize the statistical information of candidate trajectories retained after physical consistency screening to improve the characterization ability of parameter uncertainties.

[0265] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A deep learning-based runoff simulation method for rain-snow-melt mixed recharge watersheds, characterized in that, include: Acquire meteorological driving data for the target watershed; Meteorological driving data are input into the snowmelt dynamic mechanism model to obtain effective liquid water; Physical state variables are extracted using an effective liquid water-driven hydrological mechanism model, and a multi-dimensional fusion feature system is constructed accordingly. Based on preset physical consistency conditions, the sequences of the multidimensional fusion feature system are screened to obtain candidate feature trajectories; A set of features is constructed based on each candidate feature trajectory. The set of features is then input into a pre-trained time-series mapping network to output simulated runoff results. Before inputting the set features into the pre-trained temporal mapping network, there is also a pre-training process for offline construction of the deep learning model: Extract the values ​​of multiple candidate feature trajectories on the same physical state variable at each time step and calculate the mean; Calculate the sum of squares of the differences between the corresponding variable values ​​and the mean of each candidate feature trajectory, and extract the arithmetic square root of the sum of squares after dividing by the total number of trajectories to obtain the standard deviation characterizing the dispersion of the multi-trajectory distribution; The mean and standard deviation are concatenated and fused to form a set feature sequence, and the set feature sequence is truncated according to the set sliding time window scale to construct supervised training samples; The weight parameters of the initial deep learning network are iteratively updated using supervised training samples, thereby obtaining a pre-trained temporal mapping network.

2. The method according to claim 1, characterized in that, Meteorological driving data includes the total precipitation and average temperature at the current moment; Meteorological driving data are input into the snowmelt dynamic mechanism model to obtain effective liquid water, including the precipitation phase classification process: Based on the numerical relationship between the average air temperature and the preset pure snow temperature threshold and the preset pure rain temperature threshold, the total precipitation is divided into phases to obtain solid precipitation and liquid precipitation. The solid precipitation is numerically compensated using a pre-configured snowfall correction coefficient to eliminate measurement biases in solid precipitation during on-site meteorological observations.

3. The method according to claim 2, characterized in that, By inputting meteorological driving data into the snowmelt dynamic mechanism model, the effective liquid water is obtained, which also includes the liquid water-constrained refreezing process: The potential refreezing amount is calculated based on the pre-configured snow melting initiation temperature, average air temperature, pre-configured refreezing coefficient, and degree-day factor. Based on the amount of liquid precipitation and the pre-configured snow layer liquid water interception ratio coefficient, the upper limit of liquid water constraint is determined; Using the upper limit of liquid water constraint as the standard, the potential refreezing amount is calculated by minimizing the constraint to obtain the actual refreezing amount.

4. The method according to claim 1, characterized in that, In the step of extracting physical state variables and constructing a multi-dimensional fusion feature system based on them: The physical state variables were extracted as follows: snow melt, snow storage, soil moisture content, simulated total runoff, surface runoff, soil flow, and groundwater runoff; The multidimensional fusion feature system consists of eleven variables across four physical levels, including: Meteorological driving data and pre-acquired potential evapotranspiration are used to characterize externally driven factors, effective liquid water, snowmelt volume and snow storage volume are used to characterize snowmelt transition, soil moisture content is used to characterize water storage state, and simulated total runoff, surface runoff, soil flow and groundwater runoff are used to characterize source responses.

5. The method according to claim 4, characterized in that, Physical consistency conditions include multiple combinations of constraints on the system's closed state, the physical boundaries of variables, and the logic of flow generation, specifically including: Long-term water balance constraints for controlling the closed residuals of water volume in multi-year systems; Nonnegativity constraints that define the lower bounds of each state variable in a multidimensional fusion feature system; Seasonal snowmelt phase constraints based on pre-defined natural month intervals to check the peak flow periods; Source runoff conservation constraints that limit the deviation between the sum of multiple water source runoff components and the total flow; And a performance threshold constraint that controls the lower limit of historical fit goodness.

6. The method according to claim 5, characterized in that, The verification logic for screening using source runoff conservation constraints specifically includes: The summation of surface runoff, soil runoff, and groundwater runoff at the current moment yields the sum of the corresponding fast and slow flow components; Calculate the absolute residual between the sum of fast and slow flow components and the simulated total runoff, and determine that the corresponding characteristic trajectory passes the source runoff conservation constraint when the absolute residual is less than the preset deviation tolerance threshold.

7. The method according to claim 1, characterized in that, After obtaining the pre-trained temporal mapping network, the model selection also includes a multi-metric trade-off mechanism: Using historical observation data from the independent validation period, the Nash efficiency coefficient and auxiliary evaluation index of the pre-trained temporal mapping network under different training rounds are calculated. The auxiliary evaluation index includes at least the root mean square error and the mean absolute error. When the Nash efficiency coefficient deviates from the trend of auxiliary evaluation indicators, the evaluation conflict handling mechanism is triggered. Under the evaluation conflict handling mechanism, the runoff simulation error of the corresponding network state during the snowmelt sensitive period is extracted, and the network state with lower runoff simulation error is used as the discrimination criterion to complete the selection of the pre-trained time-series mapping network.

8. The method according to claim 1, characterized in that, Before filtering the sequences of the multidimensional fusion feature system based on preset physical consistency conditions, the process also includes: A random sampling strategy is used to generate multiple sets of candidate parameters for hydrological mechanism models, and physical boundary rules are set during the sampling and generation process. The physical boundary rules include: limiting the sum of the runoff generation coefficient representing subsurface runoff and the runoff generation coefficient representing interflow to be strictly less than a preset combined boundary threshold, in order to eliminate unreasonable combinations of runoff generation parameters from different sources.

9. The method according to claim 1, characterized in that, Acquiring meteorological driving data for the target watershed also includes unified data preprocessing: Missing records in the meteorological driving data are uniformly filled with zero-value boundaries, and extreme value parameters for all time periods are used to perform variable-wise normalization mapping on each variable in the meteorological driving data. During both the offline pre-training phase and the real-time simulation output phase of the model, the extreme value parameters in the normalized mapping are strictly kept consistent throughout the entire time period.

10. The method according to claim 1, characterized in that, After outputting the simulated runoff results, operational applications are also included: The simulated runoff results are then fed into the watershed water resources management system or flood early warning platform. This triggers the watershed water resources scheduling system or flood warning platform to generate hydrological scheduling strategies or flood hazard warning instructions for the snowmelt sensitive period in high-altitude and cold regions.

Images