A method for predicting evolution trend of basin runoff based on multi-source data fusion

Abstract

This invention discloses a method for predicting watershed runoff evolution trends based on multi-source data fusion, belonging to the field of hydrological and water resources prediction technology. The method includes acquiring historical hydrological and meteorological data and classifying rainfall-runoff events; calculating soil water absorption per unit time using water balance; constructing hysteresis loops by fitting wet and dry water absorption paths; clustering the hysteresis loop set to extract representative hysteresis loop families; matching the most similar hysteresis loop during prediction; inputting the data into a reinforcement learning model to determine the output path advancement position; determining the corresponding soil water absorption; and combining meteorological data to calculate the predicted runoff. This invention solves the problem of low runoff prediction accuracy caused by traditional models neglecting the differences in water absorption capacity during wet and dry processes by quantifying the nonlinear hysteresis characteristics of soil water absorption behavior.

Description

Technical Field

[0001] This invention relates to the field of hydrological and water resources prediction technology, and more specifically, to a method for predicting the evolution trend of watershed runoff based on multi-source data fusion. Background Technology

[0002] Watershed runoff refers to the horizontal flow of water that accumulates in rivers or low-lying areas after rainfall falls to the ground, is lost through soil infiltration and evaporation, and serves as an important basis for hydrological forecasting, water resource management, and flood control scheduling. Generally, rainfall, evaporation, and soil water absorption jointly determine the runoff per unit time. The actual amount of water that the soil can absorb depends on its saturation, pore structure, and prior wetting conditions.

[0003] Numerous studies have shown that even with the same current soil saturation, soil water absorption capacity can differ significantly depending on whether it is in a wetting or drying process—a phenomenon known in hydrology as the hysteresis effect. At the same saturation level, the soil rewetting curve (wetting path) and drainage curve (drying path) do not coincide. Therefore, saturation alone cannot uniquely determine soil water absorption capacity; the direction of wetting or drying must also be considered. If this hysteresis characteristic is ignored in rainfall-runoff prediction, models often misjudge the amount of water the soil can absorb, leading to overestimation or underestimation of runoff. Furthermore, different rainfall durations, intensities, and initial humidity backgrounds further alter the overall shape of the hysteresis curve, making it difficult for a single empirical model to adapt to different scenarios.

[0004] With modern watershed monitoring systems capable of simultaneously providing multi-source data such as rainfall, evaporation, soil saturation, and runoff, conditions have been created for explicitly constructing and utilizing the saturation-water absorption rate hysteresis relationship. Therefore, accurately characterizing the differences between wetting and drying paths, extracting representative hysteresis patterns, and applying them to runoff prediction are core issues that urgently need to be addressed in current hydrological simulation. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method for predicting the evolution trend of watershed runoff based on multi-source data fusion, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for predicting watershed runoff evolution trends based on multi-source data fusion includes the following steps:

[0008] Historical hydrological and meteorological data of the target watershed are acquired and divided into multiple rainfall-runoff event datasets according to the rainfall-runoff event classification rules. For each rainfall-runoff event dataset, the soil water absorption per unit time is calculated using the water balance principle. With soil saturation as the x-axis and soil water absorption per unit time as the y-axis, data points in the wetting process interval are fitted to a wetting water absorption path, and data points in the drying process interval are fitted to a drying water absorption path. The wetting and drying water absorption paths in the same rainfall-runoff event dataset are connected end-to-end in chronological order to form closed or semi-closed hysteresis loops. All hysteresis loops are collected to obtain a hysteresis loop set.

[0009] Shape feature clustering is performed on the set of hysteresis loops to extract several representative hysteresis loops to form a hysteresis loop family; in the prediction stage, based on the current soil saturation change characteristics, the most similar representative hysteresis loop is matched from the hysteresis loop family as the predicted hysteresis path.

[0010] The current measured hydrological and meteorological data and the predicted hysteresis path are input into a pre-trained reinforcement learning path advancement model. The reinforcement learning path advancement model outputs the advancement position on the predicted hysteresis path at the next moment. Based on the advancement position, the corresponding soil water absorption per unit time on the predicted hysteresis path is determined, and the predicted runoff is calculated by combining the predicted meteorological data.

[0011] Preferably, the historical hydrological and meteorological data includes the following historically measured data: rainfall sequence, evaporation sequence, soil saturation sequence, and runoff sequence; the current hydrological and meteorological data includes the following data measured within a preset time window before the current time: rainfall sequence, evaporation sequence, soil saturation sequence, and runoff sequence.

[0012] Preferably, the specific steps for dividing the historical hydrological and meteorological data into multiple rainfall-runoff event datasets according to the rainfall-runoff event classification rules are as follows:

[0013] Identify the time period from the start of rainfall to the point when runoff recedes to the baseflow level, and extract the rainfall, evaporation, soil saturation, and runoff sequences within this time period into an independent rainfall-runoff event dataset.

[0014] Preferably, the water balance principle is based on the formula Calculate the soil water absorption per unit time at each time step. ,in Rainfall per unit time Evaporation rate per unit time The unit time is the runoff volume; the difference in soil saturation between adjacent time steps is calculated, and when the difference is greater than zero, it is marked as a wet process interval, and when the difference is less than or equal to zero, it is marked as a dry process interval.

[0015] Preferably, the step of performing shape feature clustering on the hysteresis loop set to extract several representative hysteresis loops to form a hysteresis loop family includes:

[0016] Normalization is performed on each hysteresis loop in the set of hysteresis loops to map soil saturation and soil water absorption per unit time to a uniform numerical range.

[0017] The shape similarity distance matrix between different hysteresis loops is calculated using a dynamic time warping algorithm.

[0018] The K-means clustering algorithm is used to divide all hysteresis loops into K categories based on the shape similarity distance matrix, and the geometric center curve of each category is calculated. The geometric center curve is defined as the representative hysteresis loop of that category, and the K representative hysteresis loops constitute the hysteresis loop family.

[0019] Preferably, the most similar representative hysteresis loop is matched from the family of hysteresis loops as the predicted hysteresis path, specifically including:

[0020] Extract the measured soil saturation sequence within a preset time window before the current moment and construct the current change feature vector; calculate the shape similarity between the current change feature vector and the corresponding segments of each representative hysteresis loop in the hysteresis loop family; select several representative hysteresis loops with the highest shape similarity as candidate hysteresis loops;

[0021] Extract the measured soil saturation sequence within a preset time window before the current moment, and calculate the slope feature vector of the measured soil saturation sequence;

[0022] Calculate the similarity between the slope feature vector and the slope feature vector of the corresponding segment of the candidate hysteresis loop; select the candidate hysteresis loop with the smallest similarity as the predicted hysteresis path.

[0023] Preferably, the construction process of the reinforcement learning path advancement model includes:

[0024] Define a state space, where the state vector contains the following measured data within a preset time window prior to the current moment: rainfall sequence, evaporation sequence, soil saturation sequence, and runoff sequence; as well as the feature encoding of the predicted hysteresis path;

[0025] Define an action space, where the action vector is the advance position on the predicted hysteresis path, and the advance position represents the soil saturation value at the next predicted time step.

[0026] Define a reward function, which is the negative of the absolute value of the error between the predicted runoff and the actual runoff, calculated based on the actions output by the reinforcement learning path.

[0027] Preferably, the training process of the reinforcement learning path advancement model employs a deep deterministic policy gradient algorithm, including: constructing a policy network and a value network; using historical rainfall-runoff events as training samples, at each time step, the policy network outputs an action vector based on the current state vector, and the value network evaluates the Q-value of the action vector; updating the policy network parameters by minimizing the error between the predicted runoff and the actual runoff, and updating the value network parameters by minimizing the temporal difference error, until the model converges, thus obtaining a path advancement strategy that can output the optimal advancement position based on the input state.

[0028] Preferably, the specific steps for determining the corresponding soil water absorption per unit time on the predicted hysteresis path based on the advancement position are as follows:

[0029] Obtain the advancement position of the next prediction time step output by the reinforcement learning path advancement model, i.e., the soil saturation of the next prediction time step;

[0030] Determine whether the current path stage is a wet or dry process;

[0031] On the wet or dry water absorption path corresponding to the predicted hysteresis path, linear interpolation is performed using the soil saturation at the next prediction time step as an index to find the corresponding ordinate value, and this ordinate value is determined as the soil water absorption per unit time at the next prediction time step.

[0032] Preferably, the predicted runoff is calculated by combining predicted meteorological data, and the specific calculation formula is as follows:

[0033] ;

[0034] in, For the next time step, predict the runoff per unit time. To combine the unit-time predicted rainfall for the next time step obtained from weather forecasts, To combine the evaporation rate per unit time obtained from the next time step of the weather forecast, To determine the corresponding soil water absorption per unit time on the predicted hysteresis path based on the propulsion position;

[0035] When the calculation result When less than zero, Corrected to zero.

[0036] The advantages of this invention over the prior art are:

[0037] This invention takes the hysteresis effect, which is prevalent in soil water absorption behavior, as its core focus. By constructing wetting and drying paths with saturation versus water absorption rate per unit time as the coordinate axis, and calculating the actual water absorption rate based on water balance, it accurately characterizes the nonlinear properties of soil where water absorption capacity varies significantly with the wetting and drying directions at the same saturation. By splicing these two paths end-to-end into hysteresis loops in each rainfall-runoff event, and clustering all hysteresis loops by shape features to form a family of hysteresis loops composed of several representative hysteresis loops, this invention can automatically match the most suitable water absorption behavior pattern for different rainfall patterns and different initial wetting levels, making the prediction process highly physically consistent and adaptable.

[0038] Furthermore, this invention constructs a reinforcement learning path advancement model, using current saturation, rainfall, evaporation information, and the predicted hysteresis path as state inputs, and the advancement position on the hysteresis path at the next moment as the action. The optimal advancement strategy is trained through a large number of historical events. This method can automatically learn the true advancement rhythm of soil water absorption under different hydrological conditions, thereby avoiding the cumulative error caused by a fixed step size, and enabling the model to maintain predictive flexibility while adhering to the physical laws of hysteresis.

[0039] By combining hysteresis-based water absorption path selection with reinforcement learning-based propulsion mechanisms, this invention can maintain high prediction accuracy under complex rainfall conditions. Compared with traditional methods that do not consider hysteresis effects or use fixed infiltration formulas, its advantages are reflected in the following aspects: it can distinguish the differences in water absorption capacity between wet and dry directions under the same saturation conditions; it can adaptively select the most representative path that best matches the current rainfall scenario during the prediction stage; it can accurately reconstruct the time variation of future water absorption rates using dynamic propulsion mechanisms; and it significantly improves the interpretability and stability of runoff evolution trend prediction. Attached Figure Description

[0040] Figure 1 This is an overall schematic diagram of the invention;

[0041] Figure 2 This is a schematic diagram of the hysteresis loop of the present invention;

[0042] Figure 3 This is a schematic diagram of the clustering process of the present invention;

[0043] Figure 4 This is a schematic diagram of the reinforcement learning method of the present invention. Detailed Implementation

[0044] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0045] like Figure 1The diagram illustrates the overall process of the method described in this invention. This invention provides a method for predicting watershed runoff evolution trends based on multi-source data fusion. Its core concept lies in a deep understanding of the hysteresis effect in soil physical properties. Specifically, even at the same saturation level, soil absorbs or releases water significantly differently during the water absorption and wetting process and during water loss and drying process. Traditional hydrological models often ignore this nonlinear characteristic, leading to overestimation of runoff generation in the early stages of rainfall or underestimation in the receding stages. This invention digitizes this physical law by constructing closed or semi-closed hysteresis loops with soil saturation and water absorption rate per unit time as coordinate axes. It then combines this with reinforcement learning algorithms to accurately advance various states along the complex hysteresis path, thereby achieving high-precision prediction of watershed runoff.

[0046] The method involved in this invention mainly includes four stages: data infrastructure construction, physical law feature extraction, similar path matching, and rolling prediction based on reinforcement learning.

[0047] In the data infrastructure construction phase, the primary task is to acquire detailed historical hydrological and meteorological data for the target watershed. This data forms the cornerstone for the model to learn the watershed's runoff generation and concentration mechanisms. Historical hydrological and meteorological data specifically includes, but is not limited to, historically measured rainfall sequences, evaporation sequences, soil saturation sequences, and runoff sequences. Rainfall data can be obtained through inversion from rain gauges or meteorological radar deployed within the watershed; evaporation data is typically obtained using pan evaporation observations or calculated based on the Penman equation combined with meteorological parameters such as temperature and wind speed; soil saturation data can be acquired through a network of soil moisture sensors buried at different depths, and the watershed average saturation can be calculated using a weighted average method, or inverted using microwave remote sensing satellite data; runoff data originates from measured records at hydrological stations at the watershed outlet.

[0048] To extract independent physical processes from long-sequence historical data, this invention establishes strict rules for classifying rainfall-runoff events. This is because the hysteresis characteristics of soil are typically fully reflected in a complete cycle from the start of rainfall to the end of runoff. The specific classification steps are as follows:

[0049] First, a rainfall threshold is set, for example, setting a daily rainfall of more than 5 mm as the start of rainfall. Simultaneously, a runoff recession threshold is set to determine whether the runoff process has ended and returned to the baseflow level. The runoff recession threshold can be determined based on the baseflow level before the start of rainfall. For example, the average flow rate over several time steps before rainfall can be used as the baseflow level, with a fluctuation range of 1% to 5% added above this baseflow level. The runoff recession threshold lies within this fluctuation range. When the measured runoff rate drops to this threshold and remains stable over several consecutive time steps, the runoff process can be considered to have ended. The system automatically scans the time series, identifying the period from the start of rainfall, through the flood peak, until the runoff recedes to the baseflow level. The corresponding rainfall, evaporation, soil saturation, and runoff sequences within this period are extracted using data extraction techniques and combined into an independent rainfall-runoff event dataset. In this way, a continuous hydrological process can be decoupled into multiple independent physical event samples, providing standardized input for subsequent analysis.

[0050] In the physical feature extraction stage, this invention reveals the water absorption behavior within the soil based on the water balance principle. For each segmented rainfall-runoff event dataset, the soil water absorption per unit time at each time step is calculated using the water balance formula. The formula is set as follows:

[0051] ;

[0052] In the formula, This represents the amount of water absorbed by the soil per unit time. Represents rainfall per unit time. Represents the amount of evaporation per unit time. This represents runoff per unit time. The reason for using this formula is that, at the watershed scale, of the rainfall input, besides a portion evaporating back to the atmosphere and a portion flowing out of the watershed as surface runoff, the remaining portion is mainly intercepted by the soil or infiltrated; this portion is defined as soil water absorption. Since soil water absorption is difficult to measure directly on a large scale, calculating it through the residuals of the water balance is a more scientific method that conforms to the law of conservation of mass.

[0053] like Figure 2As shown in a further embodiment, after calculating the water absorption sequence, the system constructs a two-dimensional feature space with soil saturation as the abscissa and soil water absorption per unit time as the ordinate. In this space, each data point represents the soil's water absorption capacity under a specific humidity level. To capture the hysteresis effect, the data points need to be labeled. The system calculates the difference in soil saturation between adjacent time steps. When the difference is greater than zero, it indicates an increase in soil moisture, and is labeled as a wetting process interval; when the difference is less than or equal to zero, it indicates a decrease in soil moisture or no change, and is labeled as a drying process interval. Subsequently, polynomial fitting or local weighted regression smoothing techniques are used to fit the data points of the wetting process interval into a smooth wetting water absorption path, and the data points of the drying process interval into a drying water absorption path. Connecting these two paths in the same event end-to-end according to their chronological order forms a closed or semi-closed loop curve, i.e., a hysteresis loop. This hysteresis loop visually depicts how the soil changes its water absorption behavior as saturation rises and falls during this rainfall-runoff event. The size and shape of the loop directly reflect the nonlinearity of the soil pore structure and hydraulic conductivity characteristics in the watershed.

[0054] After compiling hysteresis loops from all historical events, a vast set of hysteresis loops is formed. To summarize a general response pattern for the watershed, such as... Figure 3 As shown, the present invention performs shape feature clustering on this set.

[0055] First, considering the significant differences in rainfall magnitude across different events, directly comparing values ​​could lead to the neglect of minor rainfall events. Therefore, each hysteresis loop needs to be normalized, mapping soil saturation and soil water absorption per unit time to a uniform numerical range of zero to one, ensuring comparability. Next, a dynamic time warping algorithm is used to calculate the shape similarity distance matrix between different hysteresis loops. The reason for choosing dynamic time warping over Euclidean distance is that the duration of different rainfall events varies, and this algorithm, by stretching or shortening the time axis, can find the optimal alignment between two sequences, thus more accurately measuring the similarity of curve shapes. Subsequently, the K-means clustering algorithm is used to divide all hysteresis loops into K categories based on the shape similarity distance matrix. The value of K can be determined using the elbow rule, typically between five and ten. For each category, the geometric center curve of all curves within that category is calculated, and this geometric center curve is defined as the representative hysteresis loop of that category. Finally, the K representative hysteresis loops constitute a family of hysteresis loops that can cover the main hydrological response patterns of the watershed.

[0056] Once in the prediction phase, the first step is similar path matching. The system needs to determine which historical pattern the current soil condition best matches. This involves two levels of filtering.

[0057] The first level is coarse screening, which extracts the measured soil saturation sequence within a preset time window before the current moment, such as the past 24 hours, to construct a current change feature vector. The shape similarity of this vector with the corresponding segments of each representative hysteresis loop in the hysteresis loop family is calculated. Here, the corresponding segment refers to the part of the representative hysteresis loop that overlaps with the current saturation range. Several loops with the highest shape similarity, such as the top three representative hysteresis loops, are selected as candidate hysteresis loops.

[0058] The second level is fine screening, aimed at capturing the consistency of changing trends. The system extracts the measured soil saturation sequence within a preset time window before the current moment, calculates the first difference or derivative of this sequence, and obtains the slope feature vector of the measured soil saturation sequence. This vector reflects the current rate of soil drying or wetting. Subsequently, the similarity between this slope feature vector and the slope feature vector of the corresponding segment of the candidate hysteresis loop is calculated. In this step, to ensure that the predicted dynamic trend best matches reality, the candidate hysteresis loop with the smallest feature difference, i.e., the highest similarity, is selected as the final predicted hysteresis path. This ensures that the selected path not only matches the numerical range but also highly matches the current reality in terms of the rate and direction of change.

[0059] like Figure 4 As shown, after determining the predicted path, i.e., the physical constraint trajectory, this invention introduces artificial intelligence technology for precise path advancement. This is because, although the path is determined, the specific point on the path at the next moment is affected by the nonlinearity of various meteorological factors, which simple interpolation cannot solve. Therefore, the current measured hydrological and meteorological data and the predicted hysteresis path are input into a pre-trained reinforcement learning path advancement model.

[0060] In constructing the reinforcement learning path advancement model, the state space is first defined. The state vector contains extremely rich information: the measured rainfall sequence, evaporation sequence, soil saturation sequence, and runoff sequence within a preset time window before the current moment, as well as the feature encoding of the predicted hysteresis path extracted through encoding techniques. This state definition ensures that the agent knows not only what happened in the past but also what the future physical constraint trajectory is.

[0061] Next, the action space is defined, and the action vector is designed as the advancement position on the predicted hysteresis path. This advancement position numerically represents the soil saturation value at the next prediction time step. This means that the agent's decision directly determines the evolution state of the system on the hysteresis loop.

[0062] The reward function is redefined. To guide the model to output predictions that align with physical reality, the negative value of the absolute error between the predicted runoff and the actual runoff, calculated based on the model's output actions, is used as the reward. The smaller the error, the larger the reward, prompting the agent to continuously optimize its strategy.

[0063] The training process of this model employs a deep deterministic policy gradient algorithm. This algorithm belongs to a deep reinforcement learning method based on an actor-judge architecture, and is particularly suitable for handling problems in continuous action spaces. Specifically, it involves constructing a policy network and a value network. The policy network typically uses a multilayer perceptron or recurrent neural network structure, responsible for outputting a specific action vector, i.e., the predicted saturation value, based on the current state vector. The value network also uses a deep neural network, responsible for evaluating the quality of the action output by the policy network, i.e., outputting a Q-value. During training, historical rainfall-runoff events are used as training samples. At each time step, the policy network outputs an action, and the value network scores it. The gradient descent method is used to update the policy network parameters by minimizing the error between the predicted and actual runoff, making its output more accurate; simultaneously, the time-series difference error is minimized to update the value network parameters, making its scoring more accurate. To increase the stability of training, an experience replay mechanism and a target network soft update technique are also introduced. Training continues until the model converges, ultimately resulting in an intelligent path advancement strategy that can accurately output the optimal advancement position based on complex input states.

[0064] In practical application of this strategy for prediction, the soil water absorption per unit time is determined on the predicted hysteresis path based on the advancement position output by the reinforcement learning model. The specific steps are as follows: First, obtain the soil saturation value for the next prediction time step output by the model. Then, determine whether the current path stage belongs to a wetting or drying process, based on the trend of saturation change between the current and previous moments. If, on the wetting or drying water absorption path corresponding to the predicted hysteresis path, linear interpolation is performed using the soil saturation at the next prediction time step as an index to find the corresponding ordinate value; this ordinate value is the soil water absorption per unit time for the next prediction time step. In this way, the predicted water absorption capacity considers both real-time meteorological drivers and is strictly constrained by historical physical laws, i.e., the hysteresis loop.

[0065] Finally, the predicted runoff is calculated by combining the predicted meteorological data. The specific calculation formula is as follows:

[0066] ;

[0067] in, For the next time step, predict the runoff per unit time. To combine the unit-time predicted rainfall for the next time step obtained from weather forecasts, To combine the evaporation rate per unit time obtained from the next time step of the weather forecast, This represents the soil water absorption per unit time at the next time step predicted based on the above steps; furthermore, considering physical realities, when the calculation results... When the value is less than zero, it indicates that rainfall and soil moisture are insufficient to generate runoff, at which point... Corrected to zero.

[0068] For obtaining future rainfall and evaporation data, forecast products provided by meteorological departments can be directly accessed, including short-term rainfall forecasts, hourly rainfall forecasts, future evaporation data, or meteorological elements used to calculate evaporation, such as temperature, humidity, wind speed, and radiation. These forecast data are generated by mature numerical weather prediction models and can provide reliable future rainfall and evaporation information for this invention.

[0069] In addition, when it is necessary to improve local accuracy or when meteorological forecast coverage is insufficient in some areas, machine learning prediction models for rainfall and evaporation can be constructed using historical meteorological observation data accumulated over a long period in the basin. By collecting historical rainfall, evaporation, and related meteorological elements, regression models or time series prediction models are trained to learn the mapping relationship between meteorological elements and rainfall and evaporation.

[0070] In summary, this invention fully utilizes multi-source hydrological and meteorological monitoring data, and overcomes the limitation of traditional models in distinguishing water absorption capacity during wetting and drying processes by explicitly constructing a hysteresis structure between soil saturation and soil water absorption rate per unit time. Representative water absorption patterns are extracted through hysteresis loop clustering, and the hysteresis path best matching the current wetting state is matched during the prediction stage, ensuring that runoff calculations always proceed along a physical path consistent with actual soil processes. Simultaneously, a reinforcement learning path advancement mechanism is introduced to automatically learn the optimal advancement rhythm under different rainfall conditions, improving the adaptability and stability of hourly runoff prediction. Combined with future rainfall and evaporation information generated by meteorological forecasts or machine learning, this invention maintains good prediction accuracy and interpretability under complex meteorological conditions, providing reliable technical support for watershed flood early warning, water resource allocation, and regional hydrological simulation.

[0071] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting watershed runoff evolution trends based on multi-source data fusion, characterized in that, Includes the following steps: Historical hydrological and meteorological data of the target watershed are acquired and divided into multiple rainfall-runoff event datasets according to the rainfall-runoff event classification rules. For each rainfall-runoff event dataset, the soil water absorption per unit time is calculated using the water balance principle. With soil saturation as the x-axis and soil water absorption per unit time as the y-axis, the data points of the wetting process interval are fitted to the wetting water absorption path, and the data points of the drying process interval are fitted to the drying water absorption path. Connect the wet and dry water absorption paths in the same rainfall-runoff event dataset in chronological order to form closed or semi-closed hysteresis loops; collect all hysteresis loops to obtain a hysteresis loop set. Shape feature clustering is performed on the set of hysteresis loops to extract several representative hysteresis loops to form a hysteresis loop family; in the prediction stage, based on the current soil saturation change characteristics, the most similar representative hysteresis loop is matched from the hysteresis loop family as the predicted hysteresis path. The current measured hydrological and meteorological data and the predicted hysteresis path are input into a pre-trained reinforcement learning path advancement model. The reinforcement learning path advancement model outputs the advancement position on the predicted hysteresis path at the next moment. Based on the advancement position, the corresponding soil water absorption per unit time on the predicted hysteresis path is determined, and the predicted runoff is calculated by combining the predicted meteorological data. The construction process of the reinforcement learning path advancement model includes: Define a state space, where the state vector contains the following measured data within a preset time window prior to the current moment: rainfall sequence, evaporation sequence, soil saturation sequence, and runoff sequence; as well as the feature encoding of the predicted hysteresis path; Define an action space, where the action vector is the advance position on the predicted hysteresis path, and the advance position represents the soil saturation value at the next predicted time step. Define a reward function, which is the negative of the absolute value of the error between the predicted runoff and the actual runoff, calculated based on the actions output by the reinforcement learning path.

2. The method according to claim 1, characterized in that, The historical hydrological and meteorological data includes the following historically measured data: rainfall sequence, evaporation sequence, soil saturation sequence, and runoff sequence; the current hydrological and meteorological data includes the following data measured within a preset time window before the current moment: rainfall sequence, evaporation sequence, soil saturation sequence, and runoff sequence.

3. The method according to claim 1, characterized in that, The specific steps for dividing the historical hydrological and meteorological data into multiple rainfall-runoff event datasets according to the rainfall-runoff event classification rules are as follows: Identify the time period from the start of rainfall to the point when runoff recedes to the pre-rainfall baseflow level, and extract the rainfall, evaporation, soil saturation, and runoff sequences within this time period into an independent rainfall-runoff event dataset.

4. The method according to claim 1, characterized in that, The water balance principle is based on the formula Calculate the soil water absorption per unit time at each time step. ,in Rainfall per unit time Evaporation rate per unit time The unit time is the runoff volume; the difference in soil saturation between adjacent time steps is calculated, and when the difference is greater than zero, it is marked as a wet process interval, and when the difference is less than or equal to zero, it is marked as a dry process interval.

5. The method according to claim 1, characterized in that, The step of performing shape feature clustering on the hysteresis loop set to extract several representative hysteresis loops to form a hysteresis loop family includes: Normalization is performed on each hysteresis loop in the set of hysteresis loops to map soil saturation and soil water absorption per unit time to a uniform numerical range. The shape similarity distance matrix between different hysteresis loops is calculated using a dynamic time warping algorithm. The K-means clustering algorithm is used to divide all hysteresis loops into K categories based on the shape similarity distance matrix, and the geometric center curve of each category is calculated. The geometric center curve is defined as the representative hysteresis loop of that category, and the K representative hysteresis loops constitute the hysteresis loop family.

6. The method according to claim 1 or 5, characterized in that, The process of matching the most similar representative hysteresis loop from the family of hysteresis loops as the predicted hysteresis path specifically includes: Extract the measured soil saturation sequence within a preset time window before the current moment and construct the current change feature vector; calculate the shape similarity between the current change feature vector and the corresponding segments of each representative hysteresis loop in the hysteresis loop family; select several representative hysteresis loops with the highest shape similarity as candidate hysteresis loops; Extract the measured soil saturation sequence within a preset time window before the current moment, and calculate the slope feature vector of the measured soil saturation sequence; Calculate the similarity between the slope feature vector and the slope feature vector of the corresponding segment of the candidate hysteresis loop; select the candidate hysteresis loop with the smallest similarity as the predicted hysteresis path.

7. The method according to claim 1, characterized in that, The training process of the reinforcement learning path advancement model employs a deep deterministic policy gradient algorithm, including: constructing a policy network and a value network; using historical rainfall-runoff events as training samples, at each time step, the policy network outputs an action vector based on the current state vector, and the value network evaluates the Q-value of the action vector; updating the policy network parameters by minimizing the error between the predicted runoff and the actual runoff, and updating the value network parameters by minimizing the temporal difference error, until the model converges, resulting in a path advancement strategy that can output the optimal advancement position based on the input state.

8. The method according to claim 1, characterized in that, The specific steps for determining the corresponding soil water absorption per unit time on the predicted hysteresis path based on the advancement position are as follows: Obtain the advancement position of the next prediction time step output by the reinforcement learning path advancement model, i.e., the soil saturation of the next prediction time step; Determine whether the current path stage is a wet or dry process; On the wet or dry water absorption path corresponding to the predicted hysteresis path, linear interpolation is performed using the soil saturation at the next prediction time step as an index to find the corresponding ordinate value, and this ordinate value is determined as the soil water absorption per unit time at the next prediction time step.

9. The method according to claim 1 or 8, characterized in that, The predicted runoff volume is calculated by combining predicted meteorological data, and the specific calculation formula is as follows: ； in, For the next time step, predict the runoff per unit time. To combine the unit-time predicted rainfall for the next time step obtained from weather forecasts, To combine the evaporation rate per unit time obtained from the next time step of the weather forecast, To determine the corresponding soil water absorption per unit time on the predicted hysteresis path based on the propulsion position; When the calculation result When less than zero, Corrected to zero.

Images