Natural runoff reconstruction method and system based on CNN-LSTM and attention mechanism

Abstract

The invention discloses a natural runoff reconstruction method and system based on CNN-LSTM and an attention mechanism, and relates to the technical field of natural runoff reconstruction, and the method comprises the steps: obtaining and preprocessing hydro meteorological data, and obtaining time sequence data; constructing a runoff reconstruction model, performing logic reconstruction on the time sequence data, and outputting a runoff reconstruction result; and carrying out hydrological and physical law verification on a runoff reconstruction result, and carrying out window visualization on a display interface. The technical problems that in the prior art, a natural runoff reconstruction model is difficult to consider the normal runoff overall precision and the extreme runoff fitting effect, and the reconstruction result lacks hydrological physical law verification, so that the physical credibility is insufficient are solved, and natural runoff reconstruction which considers the normal runoff precision and the extreme runoff precision and conforms to the physical law is achieved. And the accuracy and reliability of runoff reconstruction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural runoff reconstruction, and particularly relates to a natural runoff reconstruction method and system based on a CNN-LSTM and an attention mechanism. BACKGROUND

[0002] At present, the natural runoff sequence of the third national water resources survey and evaluation only covers 1956-2016, which has been difficult to accurately represent the current water resources endowment. Under the dual effects of global climate change and high-intensity human activities, the mechanism of water cycle in the basin has undergone profound evolution. In order to accurately grasp the latest situation of water resources and reveal the long sequence evolution mechanism and driving law, it is urgent to extend the natural runoff sequence to the recent period, so as to provide solid and reliable data support for water resources comprehensive planning, engineering rational allocation and drought and flood disaster risk warning. In the field of natural runoff reconstruction, the traditional runoff reconstruction method depends on a simple statistical model, which is difficult to effectively capture the multi-scale time sequence characteristics and nonlinear correlation of hydrological process, or uses a single structure of deep learning model, which only extracts features through a fixed size convolution kernel, and cannot balance the short period and long period hydrological information, and the learning of extreme runoff samples with low proportion and large influence is insufficient, resulting in significant deviation of model accuracy in extreme event reconstruction. At the same time, the traditional method only evaluates the model performance through statistical indicators, lacks the verification link based on hydrological physical law, so that the reconstruction result may violate the objective hydrological law and the physical credibility is insufficient.

[0003] The prior art natural runoff reconstruction model is difficult to balance the overall accuracy of normal runoff and the fitting effect of extreme runoff, and the reconstruction result lacks the verification of hydrological physical law, resulting in insufficient physical credibility. SUMMARY

[0004] The present application provides a natural runoff reconstruction method and system based on a CNN-LSTM and an attention mechanism, which is used to solve the technical problems in the prior art that the natural runoff reconstruction model is difficult to balance the overall accuracy of normal runoff and the fitting effect of extreme runoff, and the reconstruction result lacks the verification of hydrological physical law, resulting in insufficient physical credibility.

[0005] In view of the above problems, the present application provides a natural runoff reconstruction method and system based on a CNN-LSTM and an attention mechanism.

[0006] In a first aspect of the present application, a natural runoff reconstruction method based on a CNN-LSTM and an attention mechanism is provided, which comprises:

[0007] Hydrological and meteorological data are acquired and preprocessed to obtain time-series data. The preprocessing steps include feature engineering, sequence construction, and sample balancing. A runoff reconstruction model is constructed, and logical reconstruction is performed on the time-series data to output the runoff reconstruction results. The reconstruction steps include multi-scale convolution processing, self-attention weighting, bidirectional temporal dependency modeling, and bi-branch reconstruction fusion. The runoff reconstruction results are verified using hydrophysical laws and visualized in a window on the display interface.

[0008] A second aspect of this application provides a natural runoff reconstruction system based on CNN-LSTM and an attention mechanism, the system comprising:

[0009] The time-series data acquisition module is used to acquire hydrological and meteorological data and perform preprocessing to obtain time-series data. The preprocessing steps include feature engineering, sequence construction, and sample balancing. The runoff reconstruction result output module is used to construct a runoff reconstruction model, perform logical reconstruction on the time-series data, and output the runoff reconstruction results. The reconstruction steps include multi-scale convolution processing, self-attention weighting, bidirectional temporal dependency modeling, and bi-branch reconstruction fusion. The window visualization module is used to verify the runoff reconstruction results against hydrophysical laws and to visualize the results in a window on the display interface.

[0010] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0011] Hydrological and meteorological data are acquired and preprocessed to obtain time-series data. A runoff reconstruction model is constructed, and logical reconstruction is performed on the time-series data to output the runoff reconstruction results. The reconstruction steps include multi-scale convolution processing, self-attention weighting, bidirectional temporal dependency modeling, and bi-branch reconstruction fusion. The runoff reconstruction results are verified against hydrophysical laws and visualized in a window on the display interface. This method achieves the technical effect of reconstructing natural runoff with accuracy that takes into account both normal and extreme runoff conditions, conforming to physical laws, and improving the accuracy and reliability of runoff reconstruction. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A schematic diagram of the natural runoff reconstruction method based on CNN-LSTM and attention mechanism provided in the embodiments of this application;

[0014] Figure 2A schematic diagram of the structure of a natural runoff reconstruction system based on CNN-LSTM and attention mechanism provided in an embodiment of this application.

[0015] Figure labeling: Time series data acquisition module 10, runoff reconstruction result output module 20, window visualization module 30. Detailed Implementation

[0016] This application provides a natural runoff reconstruction method and system based on CNN-LSTM and attention mechanism to address the technical problems in existing natural runoff reconstruction models, which are unable to balance the overall accuracy of normal runoff and the fitting effect of extreme runoff, and whose reconstruction results lack verification by hydrophysical laws, resulting in insufficient physical credibility.

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] Example 1, as Figure 1 As shown, this application provides a natural runoff reconstruction method based on CNN-LSTM and attention mechanism, the method comprising:

[0019] Step S100: Obtain hydrological and meteorological data and perform preprocessing to obtain time series data. The preprocessing steps include feature engineering, sequence construction, and sample balancing.

[0020] Specifically, hydrological and meteorological data are acquired and preprocessed to obtain time-series data. The preprocessing stage includes feature engineering, sequence construction, and sample balancing. First, basic hydrological and meteorological data such as precipitation, evapotranspiration, runoff, temperature, and NDVI are collected. Feature engineering is then performed based on dimensions such as lag characteristics, rolling statistical characteristics, seasonal cycle characteristics, hydrological year characteristics, and interactive high-value characteristics. Lag values ​​of indicators such as precipitation and evapotranspiration are calculated from 1 to 24 steps to reflect the influence of historical conditions. Rolling statistical characteristics such as the mean and maximum values ​​of the 3-24 month window are generated to characterize the short-term cumulative effect. Seasonal cycle characteristics are constructed through monthly sine / cosine coding and seasonal identifiers. The hydrological year is defined as October to September of the following year, and the cumulative amount within the year is calculated to form the hydrological year characteristics. At the same time, interactive high-value characteristics such as precipitation-evapotranspiration ratio and rainfall intensity are constructed. Subsequently, the extracted features and target runoff data are standardized to eliminate dimensional differences and determine the time-series feature set. Next, sequence construction is performed, with a time step of 12 months. This means that the current runoff is predicted based on the features of the previous 12 steps, and the time-series feature set is converted into a time-series sequence with this time step as a constraint. Finally, sample balancing is carried out, and high-value runoff samples above the 85th percentile in the time-series are oversampled to solve the problem of high-value sample scarcity, and finally, time-series sequence data is generated.

[0021] Step S200: Construct a runoff reconstruction model, perform logical reconstruction on the time series data, and output the runoff reconstruction results. The reconstruction steps include multi-scale convolution processing, self-attention weighting, bidirectional temporal dependency modeling, and dual-branch reconstruction fusion.

[0022] Specifically, a runoff reconstruction model is constructed and logical reconstruction is performed on time-series data to output the runoff reconstruction results. The reconstruction process includes multi-scale convolutional processing, self-attention weighting, bidirectional temporal dependency modeling, and two-branch reconstruction fusion. First, the model reconstruction architecture is determined: it consists of multi-scale parallel one-dimensional convolutional layers, self-attention layers, bidirectional LSTM layers, and a two-branch output layer. The Bayesian optimization algorithm is used to search for optimal hyperparameters, such as convolutional kernel size and the number of LSTM units. The architecture is trained under supervision with the Nash efficiency coefficient in the high-value region as the optimization objective. At the same time, early stopping, Dropout, and L2 regularization are combined to suppress overfitting, thus completing the construction of the runoff reconstruction model. Then, logical reconstruction is performed on the time-series data: After importing the data into the model, a multi-scale parallel one-dimensional convolutional layer extracts short, medium, and long-period local features through three sets of convolutional kernels of different sizes, determined by Bayesian optimization. These features are then concatenated and normalized to obtain the convolutional feature sequence. A self-attention layer, based on the causal contribution of runoff formation, weights the features at key time points in the convolutional feature sequence, such as 1-3 days after heavy rainfall, to generate a weighted feature sequence. A bidirectional LSTM layer, with 256 units in the first layer and 128 units in the second layer, calculates the features based on "P(current... The conditional probability of "runoff (past state, future state)" is used to perform bidirectional temporal processing on the weighted feature sequence to capture bidirectional temporal dependencies and obtain a bidirectional weighted sequence. Finally, the dual-branch output layer initiates dual-branch reconstruction fusion. The main output branch uses mean squared error (MSE) as the loss function and outputs the first reconstruction result of the overall runoff based on the bidirectional weighted sequence. The high-value output branch uses Huber loss as the loss function and outputs the second reconstruction result for extreme runoff. The two types of results are dynamically weighted and fused through a gating mechanism to finally generate the runoff reconstruction result.

[0023] Step S300: Verify the runoff reconstruction results using hydrophysical laws and visualize them in the display interface.

[0024] Specifically, the process begins by deploying verification thresholds based on hydrophysical laws. The selected laws must include at least the water balance closure. This verifies the balance between runoff and water expenditures such as precipitation and evapotranspiration, as well as the rationality of the flow-water level relationship. It also verifies whether the flow and water level corresponding to the reconstructed runoff conform to actual hydrological patterns. Next, a physical reliability assessment is conducted on the runoff reconstruction results based on the verification thresholds. Abnormal reconstruction data that does not conform to the laws are eliminated to ensure the hydrological rationality of the results. Finally, the verified runoff reconstruction results, such as the reconstructed runoff time series curve and the distribution of high-value runoff events, are visualized in a window on the display interface to intuitively show the reconstruction effect.

[0025] In one possible implementation, step S100 further includes:

[0026] Step S110: Based on the feature dimensions, extract features from the hydrological and meteorological data, perform standardization processing of the features and target runoff data, and determine the time series feature set. The feature dimensions include at least lag features, rolling statistical features, seasonal cycle features, hydrological year features, and interactive high value features.

[0027] Step S120: Based on the time series feature set, perform time step-based sequence construction and high-value oversampling processing to generate the time series sequence data.

[0028] Specifically, feature extraction is performed on collected hydrometeorological data, including precipitation, evapotranspiration, runoff, temperature, and NDVI, based on five dimensions: lag characteristics, rolling statistical characteristics, seasonal cycle characteristics, hydrological year characteristics, and interactive high-value characteristics. Then, a time-series feature set is determined through standardization. In the feature extraction stage, the operations for each dimension are as follows: For the lag characteristic dimension, lag values ​​for precipitation, evapotranspiration, runoff, temperature, and NDVI at 1, 2, 3, 6, 12, and 24 steps are calculated, such as `precip_lag1` for the previous precipitation step and `runoff_lag24` for the previous 24 runoff steps, to capture the lagged impact of historical hydrological conditions on current runoff; for the rolling statistical characteristic dimension, the average values ​​for 3, 6, 12, and 24-month windows are generated to characterize the cumulative effect of short-term hydrological elements, such as the superimposed effect of continuous precipitation on runoff; and for seasonal cycle characteristics, the average values ​​for 3, 6, 12, and 24-month windows are generated to characterize the cumulative effect of short-term hydrological elements (such as the superimposed effect of continuous precipitation on runoff). The system employs a three-dimensional approach: a cyclical feature dimension, converting monthly information using sine and cosine encoding, and establishing summer and spring precipitation markers to accurately capture the seasonal variation patterns of runoff; a hydrological year feature dimension, defining October to September of the following year as a hydrological year, calculating cumulative precipitation and cumulative evapotranspiration within the year to match the natural cyclical characteristics of hydrological processes; and an interactive high-value feature dimension, constructing features such as precipitation-evapotranspiration ratio, rainfall intensity, and anterior humidity index to enhance the characterization of high-value runoff events such as floods. After feature extraction, the StandardScaler method is used to standardize all extracted features and target runoff data, eliminating dimensional differences between different indicators, and finally integrating them into a unified and standardized time-series feature set.

[0029] Based on a time-series feature set, time-series sequence data is generated through time-step-based sequence construction and high-value oversampling processing. In the sequence construction stage, referencing experience in hydrological time-series forecasting and model adaptation requirements, a time step of 12 is set. This means that the feature data from the previous 12 time points are used to predict the runoff at the current time point. Using this time step as a constraint, the feature data in the time-series feature set is divided into several consecutive time-series segments in chronological order. Each segment contains feature information from 12 time steps, forming the initial time-series sequence. In the high-value oversampling stage, considering the low proportion of high-value runoff (such as floods) samples in natural runoff data, which could lead to insufficient model learning, the distribution of target runoff data in the initial time-series sequence is first statistically analyzed. The 85th percentile is determined as the high-value judgment threshold. High-value samples with target runoff values ​​exceeding this threshold are oversampled, such as by replicating and expanding the number of high-value samples to balance the sample distribution. Finally, time-series sequence data that balances temporal continuity and sample balance is generated.

[0030] In one possible implementation, step S120 further includes:

[0031] Step S121: Set the time step, wherein the time step is the basis for predicting runoff at the next lower time node.

[0032] Step S122: Using the time step as a constraint, convert the time series feature set into a time series sequence, and perform oversampling processing on the target runoff data in the time series that are higher than the preset quantile to generate the time series sequence data.

[0033] Specifically, a time step is set and its role is defined. Combining the monthly variation patterns of natural runoff with the model's need to capture temporal dependencies, a time step of 12 is determined. The core role of this time step is to serve as the core basis for runoff prediction at lower time points. That is, by calling up hydrological and meteorological characteristic data from the previous 12 time points, such as precipitation, evapotranspiration, and lagged runoff characteristics from the previous 12 months, sufficient historical temporal information is provided to support the natural runoff prediction at the current time point, ensuring that the model can effectively learn the long-term temporal correlations in the runoff formation process.

[0034] Time series transformation and high-value oversampling are performed under the constraint of a set time step. In the time series transformation stage, with a time step of 12 as the dividing standard, the determined time series feature set, including lag features, rolling statistical features, seasonal cycle features, and other multi-dimensional features, is sliced ​​in chronological order. The continuous time series feature data is divided into several independent time series segments, each containing feature information from 12 consecutive time steps, thereby converting the structured time series feature set into an initial time series that conforms to the model input format. In the high-value oversampling stage, the overall distribution of target runoff data in the initial time series is first statistically analyzed. Considering the scarcity of high-value runoff, such as flood events, an 85th percentile is preset as the high-value judgment threshold to filter out high-value samples in the time series whose target runoff values ​​are higher than this threshold. Then, these high-value samples are oversampled by copying and expanding or synthesizing new samples to increase the number of high-value samples, thereby solving the problem of insufficient model learning due to the low proportion of high-value samples. Finally, time series data that balances temporal continuity and sample balance is generated.

[0035] In one possible implementation, step S200 further includes:

[0036] Step S210: Determine the reconstruction architecture, wherein the reconstruction architecture is constructed from multi-scale parallel one-dimensional convolutional layers, self-attention layers, bidirectional LSTM layers and dual-branch output layers.

[0037] Step S220: Based on the optimal hyperparameter search using Bayesian optimization, with the Nash efficiency coefficient in the high-value region as the optimization objective, supervised training under hyperparameter optimization is performed on the reconstruction architecture to construct the runoff reconstruction model.

[0038] Specifically, the core components of the reconstruction architecture are clearly defined, consisting of a multi-scale parallel one-dimensional convolutional layer, a self-attention layer, a bidirectional LSTM layer, and a dual-branch output layer connected sequentially. The multi-scale parallel one-dimensional convolutional layer serves as the core for feature extraction, using at least three sets of convolutional kernels of different sizes to process input data in parallel, capturing hydrological local features at short, medium, and long time scales. Each set of convolutions is stabilized through layer normalization. The self-attention layer follows the convolutional layer, calculating the autocorrelation of the feature matrix and automatically assigning higher weights to features at key time points based on the causal contribution of runoff formation, reducing redundant information interference. The bidirectional LSTM layer employs a two-layer structure, with the first layer containing 256 units and returning a complete output. The sequence preserves temporal details. The second layer has 128 units and returns the final output to compress features. Each layer is followed by a Dropout layer with a dropout rate of 0.2-0.5. Through a bidirectional structure, it captures the temporal dependencies from the past to the present and from the future to the present, adapting to the bidirectional correlation of hydrological processes. The dual-branch output layer includes a main output branch and a high-value output branch. Both adopt a fully connected layer structure of 128, 64, and 1. The main branch is used for conventional runoff prediction, and the high-value branch is specifically for high-value runoff prediction, laying the architectural foundation for subsequent differentiated optimization.

[0039] The optimal hyperparameter search is conducted based on Bayesian optimization, with the Nash efficiency coefficient (NSE) in high-value regions as the optimization objective. Supervised training under hyperparameter optimization is performed on the reconstruction architecture to ultimately construct the runoff reconstruction model. In the hyperparameter search phase, the Optuna framework is used for Bayesian optimization. The hyperparameters to be optimized include the kernel size of the multi-scale one-dimensional convolutional layers, the number of units in the bidirectional LSTM layers, the dropout rate of the Dropout layers, and the learning rate. The optimization process uses a Gaussian process model to record the training effect of historical hyperparameter combinations, using the NSE in high-value regions as the core indicator to dynamically guide the search direction of subsequent hyperparameter combinations, ensuring rapid identification of the optimal hyperparameter combination. In the supervised training phase, the generated time-series data is divided into training and validation sets, and the reconstruction architecture is configured with the optimal hyperparameter combination. Simultaneously, an early stopping mechanism is introduced. When the NSE in the high-value region of the validation set does not improve for multiple consecutive rounds, training is stopped and the learning rate decay mechanism is implemented. When the validation loss stagnates, the learning rate is reduced. Together with L2 regularization and Dropout layers, the model overfitting is suppressed. During training, the main output branch uses the mean squared error as the loss function to optimize the overall runoff prediction accuracy, while the high-value output branch uses Huber loss as the loss function and sets a loss weight of 0.2-0.5 to enhance the fitting effect of high-value runoff. This continues until the model training reaches the preset maximum number of rounds or the early stopping condition is triggered, ultimately resulting in a runoff reconstruction model with stable performance and satisfactory accuracy.

[0040] In one possible implementation, step S210 further includes:

[0041] Step S211: The one-dimensional convolutional layer performs local feature extraction and feature concatenation normalization with convolutional kernels of at least three scales, wherein the hyperparameters of the one-dimensional convolutional layer include at least the convolutional kernel size.

[0042] Step S212: The self-attention layer performs weighted processing of local features.

[0043] Step S213: The bidirectional LSTM layer performs bidirectional temporal dependency modeling of weighted local features.

[0044] Step S214: The dual-branch output layer includes a main output branch and a high-value output branch, wherein the main output branch uses mean squared error as the loss function, and the high-value output branch uses Huber loss as the loss function.

[0045] Specifically, the function and construction logic of the one-dimensional convolutional layer should be clearly defined. This layer needs to perform local feature extraction and feature concatenation normalization of convolutional kernels of at least three scales, and its hyperparameters need to be determined through hyperparameter optimization iteration. The construction process consists of two steps: The first step involves iterative hyperparameter optimization. Combining the Bayesian optimization algorithm, with the Nash efficiency coefficient (NSE) in high-value regions as the optimization target, key processing parameters of each network layer, such as the kernel size of the one-dimensional convolutional layer, the number of units in the bidirectional LSTM layer, and the overall learning rate of the model, are searched and iteratively adjusted. The training effect of historical parameter combinations is recorded through a Gaussian process model to dynamically optimize parameter selection until the optimal hyperparameter combination suitable for the time-series characteristics of hydrological data is determined. The second step, based on the optimal hyperparameter combination, initiates the feature processing flow of the one-dimensional convolutional layer. Using at least three determined convolutional kernels of different sizes adapted to short, medium, and long periods, the input time-series data is processed in parallel to capture local hydrological features at different time scales, such as short-scale daily precipitation fluctuations and long-scale weekly / monthly runoff trends. The features extracted by all convolutional kernels are spliced ​​and integrated, and then layer normalization is used to eliminate feature distribution differences to ensure the stability of subsequent module training.

[0046] Next, the self-attention layer performs weighted processing of local features. This layer is connected after the one-dimensional convolutional layer. Based on the causal contribution logic of runoff formation, it automatically identifies time node features that play a key role in the reconstruction of the current runoff by calculating the autocorrelation of the convolutional feature sequence, such as precipitation and humidity features 1 to 3 days after heavy precipitation. It assigns higher weights to these key features and reduces the weight of irrelevant noise features, such as the weight of small precipitation fluctuations that have no impact during the dry season. Through weighted processing, the guiding role of core information on the model is strengthened, and the effectiveness of feature data is improved.

[0047] Then, a bidirectional temporal dependency modeling of weighted local features is performed through a bidirectional LSTM layer. This layer adopts a two-layer structure design, with the first layer having 256 units and the second layer having 128 units. The weighted feature sequence output from the self-attention layer is used as input. The bidirectional structure captures the temporal dependency relationship between the past to the present and the future to the present simultaneously. It learns both the lagged effects of previous hydrological elements, such as the past 12 steps of precipitation and lagged runoff on the current runoff, and also considers the potential hydrological conditions in the future, such as the indirect effect of future short-term precipitation on the current watershed water storage and its correlation effect on runoff formation, adapting to the bidirectional correlation of hydrological processes. At the same time, a Dropout layer is connected after each LSTM layer to suppress model overfitting and ensure the accuracy and generalization ability of temporal dependency modeling.

[0048] Finally, the composition and loss function configuration of the dual-branch output layer are clarified. This layer includes a main output branch and a high-value output branch, both of which adopt a fully connected layer network structure of 128, 64, and 1 to match different runoff prediction needs. The main output branch uses mean squared error as the loss function, focusing on the fitting accuracy of the overall runoff sequence. By minimizing the squared error between the predicted and actual values, it ensures the accuracy of reconstructing conventional runoff. The high-value output branch uses Huber loss as the loss function. This loss function is more robust to outliers, avoiding model learning bias caused by the scarcity of high-value samples. By specifically optimizing the prediction effect in high-value regions, it solves the problem of low overall error but large bias in extreme events.

[0049] In one possible implementation, step S200 further includes:

[0050] Step S230: Import the time series data into the runoff reconstruction model, perform local feature extraction and multi-scale fusion splicing under multiple convolutional branches in the one-dimensional convolutional layer, and normalize the data to determine the convolutional feature sequence.

[0051] Step S240: Based on the causal contribution of runoff formation, the self-attention layer performs feature weighting on the convolutional feature sequence to determine the weighted feature sequence.

[0052] Step S250: The bidirectional LSTM layer performs time-series processing on the weighted feature sequence, triggering the dual-branch output layer to perform dual-path reconstruction processing and generate runoff reconstruction results.

[0053] Specifically, the generated time-series data is imported into the constructed runoff reconstruction model. A one-dimensional convolutional layer performs local feature extraction, multi-scale fusion and concatenation, and normalization under multiple convolutional branches to ultimately determine the convolutional feature sequence. This one-dimensional convolutional layer is equipped with at least three scale convolutional kernels determined by Bayesian optimization, with multiple convolutional branches operating in parallel: different scale convolutional kernels capture local features of short, medium, and long periods (e.g., 2-8 time units, 4-12 time units, and 6-16 time units) in hydrological data, respectively. The short-scale convolutional kernel extracts daily precipitation fluctuation features, while the long-scale convolutional kernel extracts weekly / monthly runoff trend features. After each branch completes local feature extraction, the features of all branches are concatenated and integrated to achieve multi-scale information fusion. Then, layer normalization is used to eliminate the distribution differences of features from different branches, avoiding the impact of feature value fluctuations on subsequent module training, and finally forming a well-structured and comprehensive convolutional feature sequence.

[0054] The core role of the self-attention layer is clearly defined: based on the causal contribution of runoff formation, it performs feature weighting on the convolutional feature sequence output by the one-dimensional convolutional layer to ultimately determine the weighted feature sequence. During the operation, the self-attention layer first analyzes the causal relationship between the features at each time point in the convolutional feature sequence and runoff formation. For example, precipitation characteristics 1-3 days after heavy rainfall and the preceding humidity characteristics during periods of high humidity have a direct and crucial driving effect on current runoff generation. Meanwhile, small precipitation fluctuations with no significant impact during the dry season and redundant features not directly related to runoff formation contribute less to runoff reconstruction. Subsequently, by calculating the autocorrelation of the feature matrix, it automatically assigns higher weights to key features with high causal contributions while reducing the weight ratio of redundant noise features with low contributions. This strengthens the guiding role of core hydrological information in subsequent time-series modeling, reduces interference from irrelevant information, and ultimately outputs a weighted feature sequence that highlights key causal features and has stronger information effectiveness, providing high-quality input for the time-series dependency modeling of the bidirectional LSTM layer.

[0055] The bidirectional LSTM layer performs temporal processing on the weighted feature sequence, which in turn triggers the dual-branch output layer to perform dual-path reconstruction processing and generate runoff reconstruction results. The bidirectional LSTM layer employs a two-layer structure: a 256-unit first layer and a 128-unit second layer. It takes the weighted feature sequence as input and captures temporal dependencies from the past to the present and from the future to the present through bidirectional modeling. It learns both past hydrological elements, such as the lagged effects of past 12 steps of runoff and accumulated precipitation on current runoff, and future potential hydrological conditions, such as the indirect effects of future short-term precipitation on watershed water storage on runoff formation. This adapts to the bidirectional correlation of hydrological processes, and the processed output is a bidirectional weighted sequence containing complete temporal dependency information. This sequence triggers a dual-branch output layer: the main output branch uses mean squared error as the loss function to output the first reconstruction result of conventional runoff based on the bidirectional weighted sequence, ensuring the overall runoff sequence fitting accuracy; the high-value output branch uses Huber loss as the loss function to output the second reconstruction result of extreme runoff based on high-value features in the bidirectional weighted sequence, enhancing the reconstruction effect of high-value events such as floods. Finally, a gating mechanism dynamically weights and fuses the two types of results to generate a runoff reconstruction result that balances overall accuracy and high-value accuracy.

[0056] In one possible implementation, step S250 further includes:

[0057] Step S251: Define the conditional probability, wherein the conditional probability is represented as P(current runoff|past state, future state).

[0058] Step S252: Based on the conditional probability, perform bidirectional time-series processing on the weighted feature sequence to determine the bidirectional weighted sequence.

[0059] Specifically, it is necessary to define a conditional probability that adapts to the temporal characteristics of hydrological processes. This conditional probability is specifically represented as "P(current runoff | past state, future state)", providing a logical basis for the temporal modeling of the bidirectional LSTM layer. The current runoff is the core target variable to be reconstructed in the model. The past state corresponds to the hydrological state information before the current time point in the weighted feature sequence, covering the lag characteristics of precipitation and rolling statistical features within the previous 12 time steps, such as accumulated precipitation. This information directly reflects the lagged driving effect of previous hydrological elements on the current runoff. The future state corresponds to the potential hydrological state information after the current time point in the weighted feature sequence, including subsequent short-term precipitation trends and evapotranspiration variation characteristics, reflecting the indirect impact of later hydrological conditions on current watershed water storage and runoff formation. This definition of conditional probability accurately matches the objective law that runoff formation in hydrological processes is affected by bidirectional temporal correlation.

[0060] Guided by the defined conditional probability P(current runoff | past state, future state), a bidirectional LSTM layer is used to perform bidirectional temporal processing on the output weighted feature sequence, ultimately determining the bidirectional weighted sequence. The bidirectional LSTM layer operates through two parallel branches: the forward branch, along the past-to-present temporal direction, performs temporal dependency learning on historical hydrological features corresponding to past states in the weighted feature sequence, such as lagged precipitation and cumulative runoff in the previous 12 steps, capturing the lagged driving patterns of previous hydrological elements on current runoff; the reverse branch, along the future-to-present temporal direction, performs temporal dependency learning on potential hydrological features corresponding to future states in the weighted feature sequence, such as subsequent short-term precipitation trend correlation features and evapotranspiration change correlation features, exploring the indirect impact of later hydrological conditions on current runoff formation. After both branches complete the temporal feature extraction, the output temporal features are fused and integrated to form a bidirectional weighted sequence that simultaneously covers both past-to-present and future-to-present temporal information, fully matching the bidirectional correlation of hydrological processes reflected by the conditional probability.

[0061] In one possible implementation, step S250 further includes:

[0062] Step S253: Based on the main output branch, perform runoff reconstruction based on the bidirectional weighted sequence, guided by the conditional expectation value of the overall runoff sequence, and determine the first reconstruction result.

[0063] Step S254: Based on the high-value output branch, perform runoff reconstruction based on the bidirectional weighted sequence, guided by the conditional distribution of extreme runoff events, and determine the second reconstruction result.

[0064] Step S255: According to the gating mechanism, perform dynamic weighted fusion of the first reconstruction result and the second reconstruction result under the optimal overall performance, and output the runoff reconstruction result.

[0065] Specifically, using the main output branch as the carrier and guided by the conditional expectation value of the overall runoff sequence, a bidirectional weighted sequence containing both past-to-present and future-to-present time-series information is taken as input to perform runoff reconstruction and determine the first reconstruction result. The main output branch adopts a fully connected layer network structure of 128, 64, and 1, coupled with a mean squared error loss function. By minimizing the squared error between the predicted and actual values, it focuses on ensuring the average reconstruction accuracy under most normal hydrological scenarios, and accurately fitting conventional runoff types such as baseflow and ordinary flood season runoff. The underlying logic of this branch design lies in the fact that normal hydrological events and extreme hydrological events, such as flood peaks, follow different generation logics and error distributions. The main output branch lays the foundation for the accuracy of the overall runoff sequence by focusing on the subtask of learning the overall conditional expectation. At the same time, it forms a task division with the high-value output branch that focuses on learning the tail conditional probability. The latter uses a more robust Huber loss function and high-value sample oversampling to specifically fit the extreme value patterns at the tail of the runoff distribution, such as flood peak identification and reconstruction. Ultimately, it achieves the goal of significantly improving the analysis capability of low-probability, high-impact flood events without significantly sacrificing overall accuracy, making the overall reconstruction results closer to the skewed and heavy-tailed distribution characteristics of the real runoff.

[0066] Using the high-value output branch as the main execution body and guided by the conditional distribution of low-probability, high-impact hydrological events such as flood peaks, the system takes a bidirectional weighted sequence, containing bidirectional time-series information from the past to the present and from the future to the present, as input to conduct targeted runoff reconstruction and determine the second reconstruction result. The high-value output branch adopts the same fully connected layer 128, 64, 1 network structure as the main output branch, but its design focuses on solving the problems of scarce extreme runoff samples and insufficient fitting of traditional models: on the one hand, relying on the oversampling of high-value samples in the early stage, the runoff samples above the 85th percentile are expanded to provide sufficient data support for the model to learn extreme runoff patterns; on the other hand, the Huber loss function is used instead of the mean squared error. This loss function is more robust to extreme values ​​and can avoid excessive interference from the error of extreme runoff samples on model training. Through the above design, the high-value output branch can accurately capture the extreme value characteristics of the tail of the runoff distribution. Based on the key information related to extreme runoff in the bidirectional weighted sequence, such as the characteristics of rainstorm intensity and the characteristics of high humidity in the early stage, it fits the conditional distribution law of extreme runoff events and finally outputs the second reconstruction result that can accurately identify and reconstruct extreme runoff events such as flood peaks.

[0067] The gating mechanism first analyzes the hydrological scene characteristics contained in the bidirectional weighted sequence, such as precipitation intensity and pre-flow humidity, to determine whether the current period is a normal or extreme runoff period. Then, it dynamically adjusts the weight ratio of the two types of results: when a normal runoff scenario is identified, the weight of the first reconstruction result is increased, such as to 0.7-0.9, prioritizing the fitting accuracy of baseflow and ordinary flood season runoff; when extreme runoff characteristics are detected, such as excessive rainfall intensity or a sudden increase in runoff, the weight of the second reconstruction result is increased, such as to 0.6-0.8, enhancing the reconstruction accuracy of extreme events such as flood peaks. This dynamic fusion approach avoids the problems of single-branch models and adapts to the different patterns of normal and extreme events through task decomposition, ultimately outputting runoff reconstruction results that more closely resemble the skewed and heavy-tailed distribution characteristics in the real world. This significantly improves the analytical capability for extreme runoff events without sacrificing overall accuracy.

[0068] In one possible implementation, step S300 further includes:

[0069] Step S310: Deploy verification thresholds according to hydrophysical laws, wherein the hydrophysical laws include at least the water balance closure degree and the rationality of the flow-water level relationship.

[0070] Step S320: Based on the verification threshold, perform a physical reliability assessment of the runoff reconstruction results and visualize the display interface.

[0071] Specifically, based on the laws of hydrophysics, a clear verification threshold is set for the physical rationality assessment of runoff reconstruction results. The selected hydrophysical laws cover at least two core dimensions: the closure of water balance and the rationality of flow-water level relationship. Specifically, when deploying the threshold for verifying the closure of water balance, based on the fundamental hydrological principle of water balance (runoff = precipitation - evapotranspiration - infiltration - water storage change), and combined with the climatic characteristics of target basins such as the Yellow River Basin, such as precipitation distribution, evapotranspiration intensity, and underlying surface conditions such as vegetation cover type and soil permeability, a quantitative allowable range for water balance closure error is set. For example, the absolute value of the closure error should not exceed 5%, ensuring that the reconstructed runoff data conforms to the objective physical laws of watershed water revenue and expenditure. When deploying the threshold for verifying the rationality of flow-water level relationship, based on the flow-water level measured data accumulated from long-term monitoring of hydrological stations within the target basin, a basin-specific flow-water level correlation model is constructed, such as the commonly used power function fitting relationship. A deviation threshold between the model's predicted value and the actual monitored value is set, such as a relative deviation not exceeding 8%, thereby defining the physical rationality standard for matching the reconstructed runoff flow data with the measured water level data.

[0072] In the physical reliability assessment stage, the final runoff reconstruction results, including both regular and extreme runoff data, are substituted into the verification thresholds for water balance closure and the rationality of the flow-water level relationship. First, the water balance closure error corresponding to the reconstructed runoff is calculated to determine if it falls within the set error range. Then, based on the flow data of the reconstructed runoff and the measured water level data for the corresponding time period, the matching relationship between the two is verified to meet the flow-water level deviation threshold requirement. Reconstruction results that simultaneously meet both threshold conditions and possess physical reliability are selected, while abnormal data that do not conform to hydrophysical laws are eliminated. In the window visualization stage, the physical reliability... The reconstruction results are presented in a multi-dimensional window format on the display interface. For example, a window for reconstructed runoff time series curves can be set up, overlaying the actual measured runoff values ​​with the reconstructed values ​​to intuitively show the fitting effect of the two over time series; a window for water balance closure error can be set up, presenting the water balance closure error for each time period in a bar chart or line chart, clearly reflecting the rationality of the water balance in the results; a window for flow-water level relationship scatter plot can be set up, marking the corresponding points of the reconstructed flow and the measured water level, overlaying the theoretical flow-water level fitting line to show the degree of matching between the two. Through multi-window visualization, users can intuitively grasp the accuracy and physical rationality of the reconstruction results.

[0073] Example 2, based on the same inventive concept as the natural runoff reconstruction method based on CNN-LSTM and attention mechanism in the previous examples, such as... Figure 2 As shown, this application provides a natural runoff reconstruction system based on CNN-LSTM and attention mechanism. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0074] The time series data acquisition module 10 is used to acquire hydrological and meteorological data and perform preprocessing to acquire time series data. The preprocessing steps include feature engineering, sequence construction and sample balancing.

[0075] The runoff reconstruction result output module 20 is used to construct a runoff reconstruction model, perform logical reconstruction on the time series data, and output the runoff reconstruction result. The reconstruction steps include multi-scale convolution processing, self-attention weighting, bidirectional temporal dependency modeling, and bi-branch reconstruction fusion.

[0076] The window visualization module 30 is used to verify the hydrophysical laws of the runoff reconstruction results and to visualize them in the display interface.

[0077] Furthermore, the system is also used to implement the following functions:

[0078] Based on the feature dimensions, feature extraction is performed on the hydrological and meteorological data, and standardization processing of the features and target runoff data is performed to determine the time series feature set. The feature dimensions include at least lag features, rolling statistical features, seasonal cycle features, hydrological year features, and interactive high value features. Based on the time series feature set, sequence construction based on time step and high value oversampling processing are performed to generate the time series data.

[0079] Furthermore, the system is also used to implement the following functions:

[0080] A time step is set, wherein the time step is used as the basis for predicting runoff at lower time nodes; with the time step as a constraint, the time series feature set is converted into a time series sequence, and the target runoff data in the time series that are higher than a preset quantile are oversampled to generate the time series sequence data.

[0081] Furthermore, the system is also used to implement the following functions:

[0082] A reconstruction architecture is determined, wherein the reconstruction architecture is constructed by multi-scale parallel one-dimensional convolutional layers, self-attention layers, bidirectional LSTM layers and dual-branch output layers; based on the optimal hyperparameter search using Bayesian optimization, with the Nash efficiency coefficient in high-value regions as the optimization objective, supervised training under hyperparameter optimization is performed on the reconstruction architecture to construct a runoff reconstruction model.

[0083] Furthermore, the system is also used to implement the following functions:

[0084] The one-dimensional convolutional layer performs local feature extraction and feature concatenation normalization using convolutional kernels of at least three scales, wherein the hyperparameters of the one-dimensional convolutional layer include at least the kernel size; the self-attention layer performs weighted processing of local features; the bidirectional LSTM layer performs bidirectional temporal dependency modeling of the weighted local features; the dual-branch output layer includes a main output branch and a high-value output branch, wherein the main output branch uses mean squared error as the loss function, and the high-value output branch uses Huber loss as the loss function.

[0085] Furthermore, the system is also used to implement the following functions:

[0086] The time-series data is imported into the runoff reconstruction model. Local feature extraction and multi-scale fusion are performed under multiple convolutional branches in the one-dimensional convolutional layer, and normalization is applied to determine the convolutional feature sequence. Based on the causal contribution of runoff formation, the self-attention layer performs feature weighting on the convolutional feature sequence to determine the weighted feature sequence. The bidirectional LSTM layer performs time-series processing on the weighted feature sequence, triggering the dual-branch output layer to perform dual-path reconstruction processing to generate the runoff reconstruction result.

[0087] Furthermore, the system is also used to implement the following functions:

[0088] Define a conditional probability, where the conditional probability is represented as P(current runoff|past state, future state); perform bidirectional time series processing on the weighted feature sequence based on the conditional probability to determine the bidirectional weighted sequence.

[0089] Furthermore, the system is also used to implement the following functions:

[0090] Based on the main output branch, runoff reconstruction based on the bidirectional weighted sequence is performed guided by the conditional expectation value of the overall runoff sequence to determine the first reconstruction result; based on the high value output branch, runoff reconstruction based on the bidirectional weighted sequence is performed guided by the conditional distribution of extreme runoff events to determine the second reconstruction result; based on the gating mechanism, dynamic weighting fusion under the optimal overall performance is performed on the first reconstruction result and the second reconstruction result to output the runoff reconstruction result.

[0091] Furthermore, the system is also used to implement the following functions:

[0092] Based on the laws of hydrophysical analysis, a verification threshold is set, wherein the laws of hydrophysical analysis include at least the water balance closure degree and the rationality of the flow-water level relationship; based on the verification threshold, the physical reliability assessment of the runoff reconstruction results is performed and the display interface is visualized.

[0093] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0094] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0095] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A natural runoff reconstruction method based on CNN-LSTM and attention mechanism, characterized in that, The method includes: Hydrological and meteorological data are acquired and preprocessed to obtain time-series data. The preprocessing steps include feature engineering, sequence construction, and sample balancing. A runoff reconstruction model is constructed, and logical reconstruction is performed on the time series data to output the runoff reconstruction results. The reconstruction steps include multi-scale convolution processing, self-attention weighting, bidirectional temporal dependency modeling and bi-branch reconstruction fusion. The runoff reconstruction results are verified using hydrophysical laws and visualized in a window on the display interface.

2. The natural runoff reconstruction method based on CNN-LSTM and attention mechanism as described in claim 1, characterized in that, Acquire and preprocess hydrological and meteorological data to obtain time-series data, including: Based on the feature dimensions, feature extraction is performed on the hydrological and meteorological data, and standardization processing of the features and target runoff data is performed to determine the time series feature set. The feature dimensions include at least lag features, rolling statistical features, seasonal cycle features, hydrological year features, and interactive high value features. Based on the time series feature set, sequence construction based on time step and high-value oversampling processing are performed to generate the time series sequence data.

3. The natural runoff reconstruction method based on CNN-LSTM and attention mechanism as described in claim 2, characterized in that, Perform time-step-based sequence construction and high-value oversampling, including: A time step is set, wherein the time step is the basis for predicting runoff at the next time node; Using the time step as a constraint, the time series feature set is converted into a time series sequence, and the target runoff data in the time series that are higher than the preset quantile are oversampled to generate the time series sequence data.

4. The natural runoff reconstruction method based on CNN-LSTM and attention mechanism as described in claim 1, characterized in that, Constructing a runoff reconstruction model includes: The reconstruction architecture is determined, wherein the reconstruction architecture is constructed by multi-scale parallel one-dimensional convolutional layers, self-attention layers, bidirectional LSTM layers and dual-branch output layers; Based on the optimal hyperparameter search using Bayesian optimization, and with the Nash efficiency coefficient in high-value regions as the optimization objective, supervised training under hyperparameter optimization is performed on the reconstruction architecture to construct a runoff reconstruction model.

5. The natural runoff reconstruction method based on CNN-LSTM and attention mechanism as described in claim 4, characterized in that, The one-dimensional convolutional layer performs local feature extraction and feature concatenation normalization with convolutional kernels of at least three scales, wherein the hyperparameters of the one-dimensional convolutional layer include at least the kernel size; The self-attention layer performs weighted processing of local features; The bidirectional LSTM layer performs bidirectional temporal dependency modeling of weighted local features; The dual-branch output layer includes a main output branch and a high-value output branch, wherein the main output branch uses mean squared error as the loss function, and the high-value output branch uses Huber loss as the loss function.

6. The natural runoff reconstruction method based on CNN-LSTM and attention mechanism as described in claim 5, characterized in that, Logical reconstruction is performed on the time-series data to output runoff reconstruction results, including: The time series data is imported into the runoff reconstruction model. Local feature extraction and multi-scale fusion splicing under multiple convolutional branches are performed in the one-dimensional convolutional layer. Normalization processing is then used to determine the convolutional feature sequence. Based on the causal contribution of runoff formation, the self-attention layer performs feature weighting on the convolutional feature sequence to determine the weighted feature sequence; The bidirectional LSTM layer performs time-series processing on the weighted feature sequence, triggering the dual-branch output layer to perform dual-path reconstruction processing and generate runoff reconstruction results.

7. The natural runoff reconstruction method based on CNN-LSTM and attention mechanism as described in claim 6, characterized in that, The bidirectional LSTM layer performs time-series processing on the weighted feature sequence, including: Define a conditional probability, where the conditional probability is represented as P(current runoff|past state, future state); Based on the conditional probability, the weighted feature sequence is subjected to bidirectional time series processing to determine the bidirectional weighted sequence.

8. The natural runoff reconstruction method based on CNN-LSTM and attention mechanism as described in claim 7, characterized in that, Triggering the dual-branch output layer to perform dual-path reconstruction processing includes: Based on the main output branch, runoff reconstruction based on the bidirectional weighted sequence is performed guided by the conditional expectation value of the overall runoff sequence to determine the first reconstruction result; Based on the high-value output branch, runoff reconstruction based on the bidirectional weighted sequence is performed, guided by the conditional distribution of extreme runoff events, to determine the second reconstruction result; According to the gating mechanism, the first reconstruction result and the second reconstruction result are dynamically weighted and fused under the condition of optimal overall performance, and the runoff reconstruction result is output.

9. The natural runoff reconstruction method based on CNN-LSTM and attention mechanism as described in claim 1, characterized in that, The runoff reconstruction results are verified using hydrophysical laws and visualized in a window on the display interface, including: Based on the laws of hydrophysics, a verification threshold is deployed, wherein the laws of hydrophysics include at least the water balance closure degree and the rationality of the flow-water level relationship; Based on the verification threshold, the physical reliability assessment of the runoff reconstruction results is performed, and the display interface is visualized.

10. A natural runoff reconstruction system based on CNN-LSTM and attention mechanism, characterized in that, The system is used to implement the natural runoff reconstruction method based on CNN-LSTM and attention mechanism as described in any one of claims 1-9, the system comprising: The time series data acquisition module is used to acquire hydrological and meteorological data and perform preprocessing to acquire time series data. The preprocessing steps include feature engineering, sequence construction and sample balancing. The runoff reconstruction result output module is used to construct a runoff reconstruction model, perform logical reconstruction on the time series data, and output the runoff reconstruction result. The reconstruction steps include multi-scale convolution processing, self-attention weighting, bidirectional temporal dependency modeling, and dual-branch reconstruction fusion. The window visualization module is used to verify the hydrophysical laws of the runoff reconstruction results and to visualize them in the display interface.

Images