Runoff prediction method and system based on hydrological state driven gating
By constructing a stable feature pool and hydrological prior features based on a hydrological state-driven gating network with a Mixture-of-Experts structure, the instability of existing runoff prediction models in non-stationary watersheds is solved, and robust runoff prediction and interpretability enhancement across regions are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-11
- Publication Date
- 2026-03-17
AI Technical Summary
Existing runoff prediction models struggle to form stable feature sets in non-stationary watersheds driven by meteorological, hydrological, and human activities. Furthermore, they lack explicit characterization of physical hydrological mechanisms such as flow patterns, regulatory behaviors, and human activities, leading to unstable predictions and insufficient interpretability.
A hydrological state-driven gating network based on the Mixture-of-Experts (MoE) structure is adopted. By constructing a stable feature pool and a hydrological prior-driven gating network, and combining multi-source data and physical prior features, expert activation weights are generated to achieve robust prediction across climate zones and control conditions.
It achieves stable prediction performance under multi-source driving and human activity interference conditions, enhances the model's cross-regional generalization ability and prediction interpretability, and is applicable to business scenarios such as runoff prediction in arid areas, water resource allocation, and urban water supply security.
Smart Images
Figure CN121682104A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of water conservancy engineering technology, and more specifically to a runoff prediction method and system based on hydrological state-driven gating. Background Technology
[0002] Currently, watershed runoff prediction schemes mainly include physical mechanism models, data-driven models, and ensemble prediction methods. Physical mechanism models (such as SWAT, WAS, and VIC) are based on water balance and energy conservation, and can describe natural hydrological processes. However, in watersheds significantly affected by reservoir regulation, water diversion projects, and social water use, they struggle to reflect the dynamic changes caused by human activities in a timely manner. Data-driven models (such as random forests, gradient boosting trees, and long short-term memory networks) rely on observational data to learn input-output relationships, but they are sensitive to input features and have limited cross-regional generalization capabilities. In recent years, multi-model fusion and gating mechanisms have been increasingly applied to hydrological prediction tasks. Some methods utilize graph structure features or local statistical features as inputs and achieve expert selection through deep networks. However, these methods typically rely on implicit feature representations driven by data, lacking explicit characterization of physical hydrological mechanisms such as flow regimes, regulatory behaviors, and human activities. Therefore, they struggle to support stable watershed runoff prediction under multi-source driven conditions and strong human activity interference.
[0003] Existing runoff prediction models still have significant shortcomings in non-stationary watersheds driven by meteorological, hydrological, and human activities. First, most existing feature selection methods are based on single-model importance or correlation assessments, lacking cross-model and cross-sample consistency verification mechanisms. This leads to fluctuations in input features under multi-source driving conditions, with changes in model structure and training samples, making it difficult to form a stable and reusable feature set. Second, existing multi-model fusion or gating methods rely heavily on statistical features, graph structure features, or implicit features generated by deep networks. Gated inputs lack hydrological priors with clear physical meaning regarding flow location, transitions between wet and dry seasons, reservoir regulation intensity, and human water use memory, making expert selection susceptible to noise interference and issues such as unstable weights and expert collapse. Furthermore, in watersheds significantly affected by reservoir regulation, water diversion projects, and multi-industry water use, runoff exhibits structural jumps and significant time-varying human regulation. Traditional models lack the ability to identify and model non-stationary features, making it difficult to adapt to flow regime changes brought about by alterations in regulation strategies, thus limiting predictive interpretability and stability. Therefore, how to provide a runoff prediction method and system based on hydrological state-driven gating is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0004] In view of this, the present invention provides a runoff prediction method and system based on hydrological state-driven gating. It introduces a stable feature pool construction mechanism and a hydrological prior-driven gating network on the basis of the Mixture-of-Experts (MoE) structure, thereby achieving robust prediction across climate zones and under different control conditions.
[0005] To achieve the above objectives, the present invention provides the following technical solution: A runoff prediction method based on hydrological state-driven gating includes the following steps: S1. Obtain multi-source data, calculate feature values, and construct an initial feature set; S2. Select core features from the initial feature set and construct a stable feature pool; S3. Further construct hydrophysical prior feature vectors based on stable feature pools; S4. Construct a gated network based on hydrological priors and generate expert activation weights; S5. Divide the expert pool into conventional machine learning models, machine learning models with time delay features, and deep time series models to obtain a set of candidate experts; S6. Calculate the gated weighted expert output as the final prediction result.
[0006] Furthermore, the calculation of eigenvalues in S1 is specifically as follows: Calculate seasonal phases: ; In the formula, , Used to indicate cyclicality within a year. For months; Calculate the drought phase: ; In the formula, Indicates a drought condition. This refers to monthly precipitation. Evaporation amount; Calculate the natural flow memory term: ; In the formula, For memory decay factor, For a moment t Natural traffic, For a moment t Natural traffic memory items, For a moment t -1 is a natural flow memory item; Water usage structure in computing industry and water intensity : ; ; In the formula, For the industry k Water consumption, For the industry j Water consumption, The natural runoff output by the physical model.
[0007] Furthermore, S2 specifically refers to: S21. Perform S resampling operations on the original training set to generate multiple data subsets: ; In the formula, For the first s The data subset generated by the second resampling For the first s Secondary sampling t Time of the first j 1 candidate feature value, For the first s Secondary sampling t The output value corresponding to the given time; S22. Train M basic learners on each resampled dataset and calculate the sample... s The model was trained to generate the first j One feature in the model m The next k The importance of each feature constitutes a four-dimensional feature importance set. I : ; ; In the formula, For the sample s The model was trained to generate the first j One feature in the model m The next k Importance of this characteristic For learning devices m Corresponding evaluation indicators k Lower features j Importance mapping function; S23. Calculate the importance consistency score of features based on the importance set. : ; In the formula, Assign importance indicator function; construct final stable feature pool: ; In the formula, To stabilize the feature pool, This is the characteristic stability threshold.
[0008] Furthermore, S3 specifically refers to: ; In the formula, For hydrophysical prior feature vectors, Hydrological state intensity, for t The quantile of the flow rate at time -1 on the FDC during the training period, where FDC is the flow rate duration curve. , As a trend priori, >0 indicates that the runoff has changed from the dry season to the wet season. <0 indicates a transition from the wet season to the dry season. for t The degree of deviation of runoff from its natural state is observed at time -1. , for t Water intensity at time -1.
[0009] Furthermore, the gated network in S4 adopts a lightweight multilayer perceptron structure, and obtains the hidden layer representation of hydrological state through nonlinear transformation. : ; In the formula, It is a nonlinear activation function used to perform nonlinear mapping on hydrological prior features. This is the weight matrix between the input layer and the hidden layer. For bias terms; Gated networks based on hidden state Generate activation weights for each expert : ; ; In the formula, This is the weight matrix from the hidden layer to the output layer. For bias terms, L For the number of experts, Indicates the first i The activation weight of each expert.
[0010] Furthermore, S5 specifically refers to: Define the set of candidate experts : ; In the formula, For conventional machine learning models, For machine learning models with time delay features, It is a deep time series model.
[0011] Furthermore, in S5, machine learning models with time-lag features add lag features to the initial feature set: ; In the formula, To add features to the feature set after lag, For the initial feature set, The maximum lag order, It is a lag operator.
[0012] Furthermore, the specific steps for calculating the gated weighted expert output in S6 are as follows: ; ; In the formula, Provided by experts For expert functions, For expert model structure, These are expert parameters.
[0013] A runoff prediction system based on hydrological state-driven gating, applying the aforementioned runoff prediction method based on hydrological state-driven gating, includes: The data acquisition module is used to acquire data from multiple sources; The feature construction module, connected to the data acquisition module, is used to build the initial feature set. The feature filtering module, connected to the feature construction module, is used to filter core features; The physical prior module, connected to the feature selection module, is used to construct hydrological physical prior feature vectors; The weight generation module, connected to the physical prior module, is used to construct the gating network and generate expert activation weights. The expert segmentation module, connected to the weight generation module, is used to segment the expert pool. The weighted calculation module, connected to the weight generation module and the expert segmentation module, is used to calculate the gated weighted expert output.
[0014] As can be seen from the above technical solution, compared with the prior art, the present invention provides a runoff prediction method and system based on hydrological state-driven gating, which has the following beneficial effects: 1. By constructing a cross-model stable feature pool and a physical prior-driven gating network, the hybrid expert model maintains stable prediction performance under different watersheds, different climate conditions and different human activity disturbance intensities, achieving cross-regional generalization capability superior to traditional static feature selection and black-box gating structures; 2. The input features are stable and the gating weights have clear physical meanings. When faced with scenarios such as changes in reservoir regulation, adjustments to the operation of water diversion projects, and changes in industry water use structure, it can still accurately identify the hydrological state and adaptively adjust the expert combination, thereby ensuring the continuity and interpretability of the prediction. 3. It can quantify the impact of human activities without relying on complete scheduling data, and can be directly applied to business scenarios such as runoff prediction in arid areas, water resource scheduling, ecological flow guarantee, and urban water supply security, with significant engineering application value and promotion potential. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0016] Figure 1 This is a flowchart of the runoff prediction method of the present invention; Figure 2 This is a schematic diagram of hydrological station observation and simulated runoff processes in an embodiment of the present invention, wherein, Figure 2 (a) Schematic diagram of runoff observation and simulation at the Meidai Hydrological Station; Figure 2 (b) is a schematic diagram of the observation and simulation of runoff processes at the Xierdaohe Hydrological Station; Figure 2 (c) is a schematic diagram of the observation and simulation of runoff processes at the Sanliang Hydrological Station; Figure 2 (d) is a schematic diagram of the observation and simulation of runoff processes at the Zhuozishan Hydrological Station; Figure 2 (e) is a schematic diagram of the observation and simulation of runoff processes at the Chenliyao Hydrological Station; Figure 2 (f) is a schematic diagram of the observation and simulation of runoff processes at the forebrain hydrological station; Figure 2 (g) is a schematic diagram of the observation and simulation of runoff process at Caijiacun Hydrological Station; Figure 2 (h) is a schematic diagram of the observation and simulation of runoff processes at the Shanghai-Nanjing hydrological station; Figure 2 (i) is a schematic diagram of the observation and simulation of runoff processes at the Sanleiba Hydrological Station; Figure 2 (j) is a schematic diagram of the observation and simulation of runoff processes at the Xiangjiaba Hydrological Station; Figure 3 This is a schematic diagram illustrating the temporal evolution of expert gate control weights and observed flow at the Caijiacun site in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the temporal evolution of expert gating weights and observed flow at the Meidai site in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the temporal evolution of expert gating weights and observed flow rates at two or three stations in an embodiment of the present invention. Figure 6 This is a schematic diagram illustrating the temporal evolution of the expert gate control weights and observed flow at the Zhuozishan site in an embodiment of the present invention. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] This invention discloses a runoff prediction method based on hydrological state-driven gating, such as... Figure 1 As shown, it includes the following steps: S1. Obtain multi-source data, calculate feature values, and construct an initial feature set; S2. Select core features from the initial feature set and construct a stable feature pool; S3. Further construct hydrophysical prior feature vectors based on stable feature pools; S4. Construct a gated network based on hydrological priors and generate expert activation weights; S5. Divide the expert pool into conventional machine learning models, machine learning models with time delay features, and deep time series models to obtain a set of candidate experts; S6. Calculate the gated weighted expert output as the final prediction result.
[0019] In this embodiment, the multi-source data acquired by S1 includes: meteorological, hydrological, and human activity data, as well as hydrological and meteorological characteristics such as precipitation. , evapotranspiration Temperature Periodic and memory variables, such as seasonal phase angles , Variables such as flow memory and drought phase; indicators of human activities, such as industry water use, water use structure, and water use intensity, expressed in terms of time duration. T The candidate feature number is used to construct the initial feature matrix: ; in, express t Time of the first j There are 10 candidate feature values.
[0020] Furthermore, the calculation of eigenvalues in S1 is specifically as follows: Calculate seasonal phases: ; In the formula, , Used to indicate cyclicality within a year. For months; seasonal phases are used to reflect the annual hydrological cycle rhythm, ensuring that the model can identify seasonal changes between wet and dry seasons; Calculate the drought phase: ; In the formula, Indicates a drought condition. This refers to monthly precipitation. Evapotranspiration; the drought phase indicates whether the month is dry or wet, helping to capture the impact of climate moisture surplus and deficit on runoff. If there is precipitation in the month... Less than evaporation ,but =1 indicates that the drought has begun; Calculate the natural flow memory term: ; In the formula, For memory decay factor, For a moment t Natural traffic, For a moment t Natural traffic memory items, For a moment t -1 is the natural flow memory term; the natural flow memory term characterizes the hysteresis inertia of the runoff process; Water usage structure in computing industry and water intensity : ; ; In the formula, For the industry k Water consumption, For the industry j Water consumption, Natural runoff output by the physical model; water use structure and water intensity Quantify the intensity of human activity in a region.
[0021] Furthermore, S2 specifically refers to: S21. Perform S resampling operations on the original training set to generate multiple data subsets: ; In the formula, For the first s The data subset generated by the second resampling For the first s Secondary sampling t Time of the firstj 1 candidate feature value, For the first s Secondary sampling t The output value corresponding to the given time; S22. Train M basic learners on each resampled dataset. In this embodiment, the learners include, but are not limited to, Random Forest, LightGBM, or XGBoost; calculate the samples. s The model was trained to generate the first j One feature in the model m The next k The importance of each feature constitutes a four-dimensional feature importance set. I : ; ; In the formula, For the sample s The model was trained to generate the first j One feature in the model m The next k Importance of this characteristic For learning devices m Corresponding evaluation indicators k Lower features j Importance mapping functions; in this embodiment, the importance mapping functions include gain, split frequency Freq, and permutation importance, etc. S23. Calculate the importance consistency score of features based on the importance set. : ; In the formula, Let be the importance indicator function. If a feature shows a stable contribution across multiple rounds of sampling, multiple models, and multiple indicators, then... The higher the value, the more stable the feature pool becomes; construct the final stable feature pool: ; In the formula, To stabilize the feature pool, This is the feature stability threshold. The final... It serves as a unified input layer for expert pool training, ensuring that different experts learn on the same stable input set, effectively improving cross-basin robustness and interpretability.
[0022] Unlike existing methods that select features based on single-model importance and / or correlation, the stable feature pool in this invention achieves consistency evaluation across models and across samples through a four-dimensional importance tensor, which is a novel statistical screening mechanism.
[0023] Furthermore, S3 specifically refers to: ; In the formula, For hydrophysical prior feature vectors, The intensity (magnitude) of the hydrological condition. for t -1 is the quantile (relative to flow regime) of the flow rate on the FDC during the training period. The FDC is the flow rate duration curve. , As a trend priori, >0 indicates that the runoff has changed from the dry season to the wet season. <0 indicates a transition from the wet season to the dry season, emphasizing the information about the turning point in runoff changes. for t The degree of deviation of runoff from its natural state is observed at time -1. , for t Water intensity at time -1.
[0024] The variables required for hydrological prior feature calculation all originate from the initial feature set and are transformed into physically meaningful state variables through a physical constraint mapping function, ensuring that the gating input and expert input are structurally independent but have the same source.
[0025] Furthermore, the gated network in S4 adopts a lightweight multilayer perceptron structure, and obtains the hidden layer representation of hydrological state through nonlinear transformation. : ; In the formula, The nonlinear activation function is used to perform nonlinear mapping on hydrological prior features. In this embodiment, the linear activation function can be the rectified linear unit ReLU function σ(x)=max(0,x), or it can be a commonly used activation function such as the Sigmoid function, Tanh function, or Leaky-ReLU. Furthermore, the ReLU function can be used to improve training stability and computational efficiency. This is the weight matrix between the input layer and the hidden layer. For bias terms; Gated networks based on hidden state Generate activation weights for each expert : ; ; In the formula, The weight matrix from the hidden layer to the output layer is designed with block parameters corresponding to the expert pool structure, so that each expert has an independent weight learning channel. For bias terms, L For the number of experts, Indicates the first i The activation weights of each expert. The softmax function ensures that the generated weights meet the requirements of non-negativity and normalization.
[0026] In this embodiment of the invention, activation weights are used. This can be interpreted as hydrological conditions. Lower Expert Conditional probability: ; This establishes an interpretable causal chain of "hydrological physical state → expert weights → expert contributions". The gating network relies only on hydrological prior vectors and does not use original flow or deep latent features as input, which makes the model universal across watersheds and years and avoids the weight drift and expert collapse problems that are prone to occur in traditional black box gating structures under non-stationary conditions.
[0027] The above structure realizes a hydrological prior-driven expert dynamic selection mechanism, providing a stable, physically consistent and interpretable weight distribution for subsequent expert weighted prediction.
[0028] Furthermore, S5 specifically refers to: Define the set of candidate experts : ; In the formula, For conventional machine learning models (such as multiple linear regression (MLR), generalized linear model (GAM), random forest (RF), etc.). For machine learning models with time delay features, their algorithm structure is similar to... The process is consistent, but a time delay feature is added to the input layer. This is for deep temporal models (such as LSTM and TCN). By simultaneously introducing regular experts, lag experts, and temporal experts into the expert pool, this embodiment enhances multi-scale response capabilities and avoids instability of deep models in small sample scenarios.
[0029] Furthermore, to enhance the model's ability to represent "short-term memory" and "hydrological inertia," models with time-delay features add lag features to the initial feature set. This effectively captures runoff persistence and rainfall-runoff delay, especially when the sample size limits the performance of deep sequence models. In S5, machine learning models with time-delay features add lag features to the initial feature set. ; In the formula, To add features to the feature set after lag, For the initial feature set, The maximum lag order, It is a lag operator.
[0030] Furthermore, the specific steps for calculating the gated weighted expert output in S6 are as follows: ; ; In the formula, Provided by experts In this embodiment, expert functions include, but are not limited to, MLR, GAM, RF, GB, LSTM, TCN, etc. For expert model structure, These are expert parameters.
[0031] In this embodiment of the invention, to prevent the gating network from collapsing into a single expert, a comprehensive loss function consisting of a prediction error term and a weight entropy penalty term is introduced to simultaneously optimize the gating parameters and the parameters of each expert. Loss Function Represented as: ; In the formula, MSE measures the deviation between the predicted value and the measured value; The weight entropy reflects the uniformity of the distribution of expert weights; For a moment t Expert weighting; Entropy regularization coefficient controls the balance of weight distribution; take If the value is greater than 0, the loss function is used to suppress weight collapse and encourage multi-expert collaboration. Based on this loss function, the gating parameters and expert parameters can be updated simultaneously through backpropagation, realizing a stable feature pool-driven prediction output and hydrological prior-driven adaptive weight selection, thereby improving the prediction reliability and model interpretability under non-stationary hydrological conditions.
[0032] In one embodiment of the present invention, the study area is the arid Dahei River Basin and the humid Yangtze River Basin. The system maintains stable predictive performance even during periods of drastic regulation and significant water use changes.
[0033] Observations and simulations of runoff processes at 10 representative hydrological stations, such as Figure 2 As shown, Figure 2 (a) Figure 2 (f) Corresponding stations in the Dahei River basin (arid zone), a tributary of the Yellow River. Figure 2 (g)- Figure 2(j) Corresponding to stations in the Yangtze River Basin (humid region), the figure compares the performance of this embodiment (SH_MOE) with single-model and multi-model ensembles. Compared with the single model, SH_MOE improves NSE by 0.06–0.24, logNSE by 0.05–0.37, and reduces the error margin by 5–30%. The gating structure of the hydrological prior embedding significantly enhances the interpretability and generalization of the model.
[0034] The temporal evolution of expert gate weights and observed flow at representative sites is as follows: Figures 3-6 As shown, the height of each colored band represents the relative contribution (weight) of a specific expert at that time, with the total weights summing to 1. The gating network's response to prior features such as climate seasonality, lagged flow, and human activity intensity reveals the model's weight adaptive capability at different hydrological stages, providing physical support for understanding the "input perturbation – output stability" mechanism. Research indicates that embedding hydrological priors into a data-driven architecture can effectively enhance the model's stability and interpretability under non-stationary conditions, providing theoretical support and methodological innovation for runoff prediction and intelligent water resource management in watersheds affected by human activities.
[0035] In this embodiment of the invention, the main algorithms and parameter tuning ranges are shown in Table 1, and the remaining training details follow the time series cross-validation and grid search rules. Table 1 Main Algorithms and Parameter Tuning Range
[0036] and Figure 1 Corresponding to the method described above, this embodiment of the invention also discloses a runoff prediction system based on hydrological state-driven gating. Applying the above-described runoff prediction method based on hydrological state-driven gating includes: The data acquisition module is used to acquire data from multiple sources; The feature construction module, connected to the data acquisition module, is used to build the initial feature set. The feature filtering module, connected to the feature construction module, is used to filter core features; The physical prior module, connected to the feature selection module, is used to construct hydrological physical prior feature vectors; The weight generation module, connected to the physical prior module, is used to construct the gating network and generate expert activation weights. The expert segmentation module, connected to the weight generation module, is used to segment the expert pool. The weighted calculation module, connected to the weight generation module and the expert segmentation module, is used to calculate the gated weighted expert output.
[0037] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.
[0038] Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A hydrologic state driven gate-based runoff prediction method, characterized in that, The method comprises the following steps: S1, acquiring multi-source data, calculating characteristic values, and constructing an initial characteristic set; S2, screening core characteristics from the initial characteristic set to construct a stable characteristic pool; S3, further constructing a hydrological physical prior characteristic vector based on the stable characteristic pool; S4, constructing a gate network based on hydrological priors to generate expert activation weights; S5, dividing the expert pool into a conventional machine learning model, a machine learning model containing time lag features, and a deep time series model to obtain a candidate expert set; S6, calculating the gate-weighted expert output as the final prediction result.
2. The runoff prediction method based on hydrological state driving gate according to claim 1, characterized in that, The calculation of characteristic values in S1 is as follows: Calculate the seasonal phase: ； In the formula, , for indicating the annual cycle, is the month; Calculate the drought phase: ； wherein represents a drought condition, is the monthly precipitation, is the evapotranspiration; Calculate the natural flow memory term: ； wherein is a memory decay factor, is the time t of the natural flow, is the time t of the natural flow memory term, is the time t -1 of the natural flow memory term; Computing industry water structure and water intensity : ； ； wherein is the water use of the industry k is the water use of the industry is the water use of the industry j is the water use of the industry is the natural runoff output by the physical model.
3. The runoff prediction method based on hydrological state driven gating according to claim 1, wherein, S2 is as follows: S21, resample the original training set S times to generate multiple data subsets: ； In the formula, is the first s resampling generated data subset, is the first s resampling t moment the first j candidate feature value, is the first s resampling t corresponding output value at the moment; S22, training M kinds of basic learners on each resampling data set, calculating samples s The first feature importance of the first feature under the model j The second feature importance of the second feature under the model m The third feature importance of the third feature under the model k The fourth feature importance of the fourth feature under the model I : ； ； In the formula, is a sample s is generated by model training j is the importance of the first m feature under the model k is the importance of the first learner m corresponding evaluation index k is the importance mapping function of the feature j under S23. Calculate the feature importance consistency score of the features according to the importance set : ； In the formula, is an importance indication function; constructing a final stable feature pool: ； In the formula, is a stable feature pool, is a feature stability threshold.
4. The runoff prediction method based on hydrological state driven gating according to claim 1, wherein, S3 is as follows: ； wherein, is the hydrological physical prior vector, is the hydrological state intensity, is the t is the quantile of the flow at time -1 on the flow duration curve, FDC, in the training period, , is the trend prior, > 0 indicates a shift from dry to wet period, < 0 indicates a shift from wet to dry period, is the t is the deviation of the observed runoff from the natural at time -1, , is the t is the water use intensity at time -1.
5. The runoff prediction method based on hydrological state driven gating according to claim 1, wherein, The gating network in S4 adopts a lightweight multi-layer perceptron structure, and obtains a hydrological state hidden layer representation through nonlinear transformation : ； In the formula, is a nonlinear activation function, used for nonlinear mapping of hydrological prior features, is a weight matrix between the input layer and the hidden layer, is a bias term; Gating networks based on hidden layer states Generating activation weights for each expert : ； ； wherein, is a weight matrix from the hidden layer to the output layer, is a bias term, L is the number of experts, denotes the activation weight of the i expert.
6. The runoff prediction method based on hydrological state driven gating according to claim 1, wherein, S5 is as follows: Defining a set of candidate experts : ； wherein is a conventional machine learning model, is a machine learning model with time-lag characteristics, is a deep time-series model.
7. The runoff prediction method based on hydrological state driven gating according to claim 1, wherein, In S5, the machine learning model containing time lag features adds lag features to the initial characteristic set: ； wherein is the feature set after adding the hysteresis feature, is the initial feature set, is the maximum hysteresis order, is the hysteresis operator.
8. The runoff prediction method based on hydrological state driven gating according to claim 1, wherein, In S6, the calculation of the gate-weighted expert output is as follows: ； ； wherein, is an expert output, is an expert function, is an expert model structure, is an expert parameter.
9. A hydrologic state driven gate-based runoff prediction system, comprising: The runoff prediction method based on hydrological state-driven gating according to any one of claims 1-8 comprises: a data acquisition module for acquiring multi-source data; a characteristic construction module connected with the data acquisition module for constructing an initial characteristic set a characteristic screening module connected with the characteristic construction module for screening core characteristics; a physical prior module connected with the characteristic screening module for constructing a hydrological physical prior characteristic vector; a weight generation module connected with the physical prior module for constructing a gate network and generating expert activation weights; an expert division module connected with the weight generation module for dividing the expert pool; a weighting calculation module connected with the weight generation module and the expert division module for calculating the gate-weighted expert output.