A hydrological forecasting method based on deep ensemble learning and physical mechanisms

Abstract

The present invention provides a hydrological forecasting method based on deep integrated learning and physical mechanisms, which belongs to the technical field of hydrological forecasting. The method collects remote sensing data and time series data, uses CNN and LSTM to extract features respectively, and fuses them through a dual cross-attention module. Then, the hydrological physical model SWAT is combined with the deep learning model CNN to obtain enhanced features. Multiple deep learning models are constructed as basic learners, trained using training set data, and the final prediction results are formed through an integration strategy. The present invention adopts the above-mentioned hydrological forecasting method based on deep integrated learning and physical mechanisms, combines the advantages of physical mechanisms and deep learning, and improves the accuracy and stability of hydrological forecasts. This method is of great significance for water resources management and flood prevention and disaster reduction.

Description

Technical Field

[0001] The present invention relates to the technical field of hydrological forecasting, and in particular to a hydrological forecasting method based on deep integrated learning and physical mechanisms. Background Art

[0002] Hydrological forecasting is a crucial tool for water resource management and flood prevention and mitigation, and its importance is self-evident. In today's world of increasing water scarcity and frequent floods, accurate hydrological forecasting is crucial for protecting people's lives and property and ensuring the rational use of water resources. However, traditional hydrological forecasting methods face numerous challenges, limiting their effectiveness in practical applications.

[0003] Traditional hydrological forecasting methods primarily rely on hydrological physics models. These models simulate and predict hydrological processes based on physical mechanisms and can, to a certain extent, reflect the operating patterns of hydrological systems. Representative models, such as the SWAT model, have been widely used in water resources management and environmental science research. These models simulate and predict watershed hydrological dynamics by simulating hydrological processes such as rainfall, runoff, and evapotranspiration. However, these models are often limited in their ability to handle complex hydrological processes by factors such as data acquisition, model parameter settings, and computational efficiency.

[0004] First, data acquisition is a major challenge in the application of hydrological models. Due to the complexity and uncertainty of hydrological systems, obtaining comprehensive and accurate hydrological data is often difficult. Furthermore, significant variations in natural environments and socioeconomic conditions across regions lead to uneven temporal and spatial distribution of hydrological data, further complicating data acquisition. Incomplete and inaccurate data directly impact the simulation performance and prediction accuracy of hydrological models.

[0005] Secondly, model parameter setting is also a significant factor influencing the prediction accuracy of hydrological and physical models. Hydrological and physical models typically contain a large number of parameters, the values ​​of which directly affect the model's simulation performance and prediction results. However, due to the complexity and uncertainty of hydrological systems, determining model parameters is often difficult. In practical applications, empirical formulas and historical data fitting are commonly used to determine model parameters. However, these methods often involve certain subjectivity and uncertainty, resulting in inaccurate and unstable model parameter values.

[0006] Furthermore, computational efficiency is another significant factor limiting the widespread application of hydrophysical models. Because hydrophysical models typically involve numerous physical processes and computational steps, they are computationally intensive and inefficient. This issue is particularly acute when simulating large-scale watersheds or complex hydrological systems. This not only increases the cost and time required for model application, but also limits the scope and effectiveness of the models in practical applications.

[0007] To address the challenges of traditional hydrological forecasting methods, machine learning algorithms such as deep learning offer new insights and approaches. Deep learning models, such as convolutional neural networks (CNNs) and long short-term memory (LSTM) neural networks, can automatically learn features and make predictions from large amounts of data, demonstrating strong nonlinear fitting and adaptive capabilities. These models excel at processing complex, nonlinear, and high-dimensional data, and have therefore garnered widespread attention and application in the field of hydrological forecasting.

[0008] However, pure deep learning models also have limitations. Due to the lack of physical constraints, deep learning models are prone to producing irrational predictions when data is sparse or abnormal. Furthermore, deep learning model predictions are often difficult to interpret and understand, which complicates their application and validation.

[0009] Therefore, how to combine the physical mechanisms of hydrological physics models and the data-driven capabilities of deep learning models to improve the accuracy and stability of hydrological forecasts has become a hot topic and difficulty in current research on hydrological forecasting technology. Summary of the Invention

[0010] The purpose of this invention is to provide a hydrological forecasting method based on deep ensemble learning and physical mechanisms. This method combines the advantages of physical mechanisms and deep learning to improve the accuracy and stability of hydrological forecasts. This method is of great significance for water resource management and flood prevention and disaster reduction.

[0011] To achieve the above objectives, the present invention provides a hydrological forecasting method based on deep ensemble learning and physical mechanisms, comprising the following steps:

[0012] S1. Collect and preprocess remote sensing data and time series data, construct a dataset, and divide the dataset into training set, validation set, and test set;

[0013] S2. On the training set and validation set, the convolutional neural network (CNN) is used to extract remote sensing data features from the remote sensing data, denoted as , using the long short-term memory neural network LSTM to extract time series features from time series data, denoted as ;

[0014] S3, yes and Perform normalization and use the double cross attention module to normalize the remote sensing features , normalized time series features Perform fusion to obtain fusion features;

[0015] S4. Combine the hydrological physics model with the deep learning model to obtain enhanced features;

[0016] S5. Build multiple deep learning models as basic learners, use the training set data to train the basic learners, and fuse the prediction results of the basic learners through an integration strategy to form the final prediction result. Then, use the validation set data to verify the model and obtain the optimized deep ensemble learning model.

[0017] S6. Input the test set data into the optimized deep ensemble learning model to generate flood forecast results.

[0018] Preferably, in step S1, the time series data includes hydrological data and meteorological data, the hydrological data includes water level, water temperature, water evaporation, flow, groundwater level, and ice conditions, the meteorological data includes temperature, precipitation, air pressure, and wind speed, and the remote sensing data includes MODIS data products and meteorological satellite data.

[0019] Preferably, in step S1, the preprocessing includes data cleaning and normalization;

[0020] The normalization formula is as follows:

[0021] ;

[0022] in, is the original pixel value or the original time series data value; is the normalized pixel value or the normalized time series data value; It is the minimum value of pixel value or the minimum value of time series data value; It is the maximum value of pixel value or the maximum value of time series data.

[0023] Preferably, in step S2, the convolutional neural network (CNN) includes a convolution layer, an activation function, and a pooling layer, and the long short-term memory (LSTM) neural network includes a forget gate, an input gate, a unit state candidate, a unit state update, an output gate, and a hidden state update;

[0024] Among them, the convolution layer:

[0025] ;

[0026] in, Is the position after the convolution operation Output; is the convolution kernel; is the input remote sensing image; Represents the convolution operation; The input image is The pixel value at ; Is the position of the convolution kernel relative to the output feature map Offset The value at 、 Is the position coordinate on the output feature map; 、 is the position coordinate on the convolution kernel;

[0027] The activation function is the ReLU activation function;

[0028] The forget gate, input gate, cell state candidate, cell state update, output gate, and hidden state update are as follows:

[0029] ;

[0030] ;

[0031] ;

[0032] ;

[0033] ;

[0034] ;

[0035] in, is the output of the forget gate; is the sigmoid function; is the weight matrix of the forget gate; represents the vector formed by concatenating the hidden state of the previous time step and the input of the current time step; is the hidden state of the previous time step; is the input of the current time step; is the bias term of the forget gate; is the output of the input gate; is the weight matrix of the input gate; is the bias term of the input gate; Candidate for unit status; is the updated unit status; is the bias term of the cell state candidate; is the updated cell state; is the cell state at the previous time step; represents element-wise multiplication; is the output of the output gate; is the weight matrix of the output gate; is the bias term of the output gate; is the updated hidden state.

[0036] Preferably, in step S3, the dual cross attention module includes two parts: channel attention and spatial attention. The channel attention calculation and Channel attention is achieved through global average pooling and fully connected layers;

[0037] Spatial Attention Computation and The spatial attention is achieved through global average pooling and fully connected layers.

[0038] Preferably, in step S3, a dual cross attention module is used to fuse the normalized remote sensing features and the normalized time series features. The specific operation is as follows:

[0039] Channel Fusion:

[0040] ;

[0041] in, is the channel fusion feature; and They are and The channel attention weight of

[0042] Spatial Fusion:

[0043] ;

[0044] in, It is the spatial fusion feature; and They are and The spatial attention weight of

[0045] Fusion features:

[0046] ;

[0047] in, It is a fusion feature;

[0048] Using principal component analysis Perform dimensionality reduction processing.

[0049] Preferably, in step S4, the hydrological physics model is a SWAT model, and the deep learning model is a CNN;

[0050] Quantify the SWAT model output:

[0051] ;

[0052] in, It is the quantitative output of the SWAT model; is the actual hydrological data; is a quantization function;

[0053] Fuse the quantized output of the SWAT model with the fusion features of step S3:

[0054] ;

[0055] in, It is the fusion feature; is the fusion function;

[0056] The fused features are used to train the CNN model, and the loss function is as follows:

[0057] ;

[0058] in, is the loss function; is the output predicted by the model; is the actual hydrological data;

[0059] Use the backpropagation algorithm to optimize the weights of the CNN model:

[0060] ;

[0061] in, is the updated weight; is the old weight; is the learning rate; is the loss function Weight The partial derivative of

[0062] Integrate the physical process constraints provided by the SWAT model into the CNN model:

[0063] ;

[0064] in, is the final prediction output, is an adjustment function that takes physical constraints into account;

[0065] Use the Monte Carlo method to estimate the conditional probability distribution of flood forecast results:

[0066] ;

[0067] in, The features after fusion The final prediction output probability; is the indicator function; is the number of simulations; Each simulation , use the CNN model to make predictions and get the prediction output; For the Parameter samples for this simulation.

[0068] Preferably, in step S5, multiple deep learning models are constructed as basic learners, and the basic learners are trained using the training set data. The specific operations are as follows:

[0069] Construct multiple base learners and record them as ;

[0070] Train the base learner:

[0071] ;

[0072] in, It is a basic learner The optimal parameters of is the input data; is the true label; is the training set data; Is to use parameters Basic learner For input data predictions.

[0073] Therefore, the present invention adopts the above-mentioned hydrological forecasting method based on deep ensemble learning and physical mechanism, and the beneficial technical effects are as follows:

[0074] The hydrological forecasting method based on deep integrated learning and physical mechanisms provided by the present invention has significant technical advantages over traditional hydrological forecasting methods.

[0075] First, by introducing deep learning models, specifically convolutional neural networks (CNNs) and long short-term memory neural networks (LSTMs), this method fully leverages the rich information in remote sensing and time series data to automatically learn and extract key features, thereby improving the accuracy of hydrological forecasts. Furthermore, by fusing remote sensing and time series features through a dual cross-attention module, the model's ability to capture complex hydrological processes is further enhanced.

[0076] Secondly, this invention combines hydrological physics models (such as the SWAT model) with deep learning models. By quantifying the SWAT model output and fusing it with fusion features, it achieves a complementary effect between physical mechanisms and data-driven capabilities. This not only improves the rationality of forecast results but also enhances the model's robustness in situations with sparse or abnormal data.

[0077] Furthermore, the present invention adopts a deep ensemble learning strategy, constructing multiple deep learning models as basic learners and fusing their prediction results through an ensemble strategy. This strategy can effectively reduce the prediction risk of a single model and improve the stability and reliability of the overall forecast results.

[0078] In summary, the hydrological forecasting method based on deep integrated learning and physical mechanisms provided by the present invention has significant beneficial technical effects in improving forecast accuracy, enhancing model robustness, and improving the stability of forecast results, and provides new ideas and methods for the further development of hydrological forecasting technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 This is a flow chart of a hydrological forecasting method based on deep integrated learning and physical mechanisms of the present invention;

[0080] Figure 2 Flowchart for remote sensing and time series data feature fusion;

[0081] Figure 3 This is the architecture diagram of the deep integrated learning model. DETAILED DESCRIPTION

[0082] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0083] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0084] Example 1

[0085] like Figure 1 FIG. 1 is a flow chart of a hydrological forecasting method based on deep ensemble learning and physical mechanism according to the present invention, which specifically includes the following steps:

[0086] Step S1: collect remote sensing data and time series data and preprocess them to construct a dataset, and divide the dataset into a training set, a validation set, and a test set.

[0087] Data source overview:

[0088] Remote sensing data: MODIS data products, high-resolution satellite data, and meteorological satellite data will be collected to ensure data diversity and integrity, covering the target watershed and surrounding areas. This data will be used to extract key information such as surface water changes, vegetation cover, and soil types.

[0089] Time series data: This includes hydrological data such as water level, water temperature, evaporation, flow, groundwater level, and ice conditions, as well as meteorological data such as temperature, precipitation, air pressure, and wind speed. These data will be used to capture the dynamic changes in hydrological and meteorological processes within the basin.

[0090] Data preprocessing:

[0091] Data cleaning: Data cleaning techniques such as interpolation, outlier detection and replacement are used to ensure the accuracy and integrity of the data.

[0092] Normalization: Normalize the collected data so that its value is in the range of [0,1] to improve model training efficiency and prediction accuracy.

[0093] The normalization formula is as follows:

[0094] ;

[0095] in, is the original pixel value or the original time series data value; is the normalized pixel value or the normalized time series data value; It is the minimum value of pixel value or the minimum value of time series data value; It is the maximum value of pixel value or the maximum value of time series data.

[0096] Dataset partitioning: Divide the preprocessed dataset into training set, validation set, and test set to ensure that the model has sufficient data support during the training, validation, and testing stages.

[0097] Step S2: On the training set and the validation set, the convolutional neural network (CNN) is used to extract remote sensing data features from the remote sensing data, which is recorded as , using the long short-term memory neural network LSTM to extract time series features from time series data, denoted as .

[0098] CNN feature extraction: Convolutional neural networks (CNNs) are used to extract spatial distribution features such as surface water changes, vegetation cover, and soil types from remote sensing data. Feature extraction is optimized by adjusting the CNN architecture and parameters.

[0099] LSTM feature extraction: Long short-term memory (LSTM) neural networks are used to extract dynamic features such as water level changes, flow changes, and temperature fluctuations from time series data. The LSTM's recurrent structure enables it to capture long-term dependencies in time series data.

[0100] Step S3: and Perform normalization and use the double cross attention module to normalize the remote sensing features , normalized time series features Perform fusion to obtain fusion features.

[0101] Normalization: The features extracted by CNN and LSTM are normalized again to ensure the consistency of the dimensions of the features.

[0102] Dual Cross Attention Module: This module is introduced to fuse normalized remote sensing features and time series features. This module achieves effective fusion and complementarity between features by calculating channel attention and spatial attention.

[0103] The dual cross attention module includes two parts: channel attention and spatial attention. Channel attention calculation and The channel attention is achieved through global average pooling and fully connected layers.

[0104] Spatial Attention Computation and The spatial attention is achieved through global average pooling and fully connected layers.

[0105] like Figure 2 The figure shows the flow chart of remote sensing and time series data feature fusion.

[0106] Channel Fusion:

[0107] ;

[0108] in, is the channel fusion feature; and They are and The channel attention weights of .

[0109] Spatial Fusion:

[0110] ;

[0111] in, It is the spatial fusion feature; and They are and The spatial attention weights.

[0112] Fusion features:

[0113] ;

[0114] in, It is a fusion feature;

[0115] Dimensionality reduction: Use principal component analysis (PCA) and other dimensionality reduction techniques to reduce the fusion features and reduce the complexity and computational complexity of the model.

[0116] Step S4: Combine the hydrological physics model with the deep learning model to obtain enhanced features.

[0117] SWAT model application: Run the hydrological physics model SWAT to simulate the hydrological process in the basin and obtain hydrological physics parameters such as soil moisture and surface runoff.

[0118] Feature Fusion: The quantized output of the SWAT model is integrated with the fused features to achieve the complementarity of physical mechanisms and data-driven capabilities. By quantizing the output of the SWAT model, it is converted into a format compatible with the fused features for subsequent model training and prediction.

[0119] Specifically:

[0120] Quantify the SWAT model output:

[0121] ;

[0122] in, It is the quantitative output of the SWAT model; is the actual hydrological data; is a quantization function.

[0123] Fuse the quantized output of the SWAT model with the fusion features of step S3:

[0124] ;

[0125] in, It is the fusion feature; is the fusion function.

[0126] The fused features are used to train the CNN model, and the loss function is as follows:

[0127] ;

[0128] in, is the loss function; is the output predicted by the model; It is the actual hydrological data.

[0129] Use the backpropagation algorithm to optimize the weights of the CNN model:

[0130] ;

[0131] in, is the updated weight; is the old weight; is the learning rate; is the partial derivative of the loss function with respect to the weights.

[0132] Integrate the physical process constraints provided by the SWAT model into the CNN model:

[0133] ;

[0134] in, is the final prediction output, is an adjustment function that takes physical constraints into account.

[0135] Use the Monte Carlo method to estimate the conditional probability distribution of flood forecast results:

[0136] ;

[0137] in, The features after fusion The final prediction output probability; is the indicator function; is the number of simulations; Each simulation , use the CNN model to make predictions and get the prediction output; For the Parameter samples for this simulation.

[0138] Step S5: deep ensemble learning model construction and training.

[0139] Model construction: Build multiple deep learning models as basic learners, including CNN and LSTM models with different architectures. Optimize model performance by adjusting the model architecture, parameters, and training strategies.

[0140] like Figure 3 As shown in the figure, it is a deep ensemble learning model architecture diagram, specifically: build multiple basic learners and record them as .

[0141] Train the base learner:

[0142] ;

[0143] in, It is a basic learner The optimal parameters of is the input data; is the true label; is the training set data; Is to use parameters Basic learner For input data predictions.

[0144] Model training: Each base learner is trained using the training set data, and the prediction results of the base learners are combined using an ensemble strategy (such as weighted averaging or voting). The choice of ensemble strategy should be determined based on the characteristics and requirements of the specific problem.

[0145] Model Validation: Use validation data to validate the model and evaluate its performance. Optimize the model's prediction accuracy and stability by adjusting model parameters and integration strategies.

[0146] Step S6: Model verification and prediction.

[0147] Test set prediction: The test set data is input into the optimized deep ensemble learning model to generate flood forecast results. During the prediction process, the accuracy and completeness of the input data should be ensured to improve the reliability of the prediction results.

[0148] Results Analysis: Compare the model prediction results with the actual hydrological data and calculate indicators such as prediction accuracy, mean square error (MSE), and mean absolute error (MAE). By analyzing these indicators, we can evaluate the model's prediction performance and identify possible areas for improvement.

[0149] It is worth noting that the contents not elaborated in detail in the present invention are all prior art and are well known to those skilled in the art.

[0150] Therefore, the present invention adopts the aforementioned hydrological forecasting method based on deep ensemble learning and physical mechanisms, combining the advantages of physical mechanisms and deep learning to improve the accuracy and stability of hydrological forecasts. This method is of great significance for water resource management and flood prevention and disaster reduction.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A hydrological forecasting method based on deep ensemble learning and physical mechanisms, characterized by: The following steps are involved: S1. Collect and preprocess remote sensing data and time series data, build a dataset, and divide the dataset into training set, validation set, and test set; S2. On the training set and validation set, the convolutional neural network (CNN) is used to extract remote sensing data features from the remote sensing data, denoted as , using the long short-term memory neural network LSTM to extract time series features from time series data, denoted as ; S3, yes and Perform normalization and use the double cross attention module to normalize the remote sensing features , normalized time series features Perform fusion to obtain fusion features; S4. Combine the hydrological physics model with the deep learning model to obtain enhanced features; S5. Build multiple deep learning models as basic learners, use the training set data to train the basic learners, and fuse the prediction results of the basic learners through an integration strategy to form the final prediction result. Then, use the validation set data to verify the model and obtain the optimized deep ensemble learning model. S6. Input the test set data into the optimized deep ensemble learning model to generate flood forecast results; In step S1, the time series data includes hydrological data and meteorological data. The hydrological data includes surface water level, water temperature, water evaporation, flow, groundwater level, and ice conditions. The meteorological data includes temperature, precipitation, air pressure, and wind speed. The remote sensing data includes MODIS data products and meteorological satellite data. In step S3, the dual cross attention module includes two parts: channel attention and spatial attention. The channel attention calculation and Channel attention is achieved through global average pooling and fully connected layers; Spatial Attention Computation and Spatial attention is achieved through global average pooling and fully connected layers; In step S3, a dual cross attention module is used to fuse the normalized remote sensing features and the normalized time series features. The specific operations are as follows: Channel Fusion: ； in, is the channel fusion feature; and They are and The channel attention weight of Spatial Fusion: ； in, It is the spatial fusion feature; and They are and The spatial attention weight of Fusion features: ； in, It is a fusion feature; Using principal component analysis Perform dimensionality reduction processing; In step S4, the hydrological physics model is the SWAT model and the deep learning model is the CNN; Quantify the SWAT model output: ； in, It is the quantitative output of the SWAT model; is the actual hydrological data; is a quantization function; Fuse the quantized output of the SWAT model with the fusion features of step S3: ； in, It is the fusion feature; is the fusion function; The fused features are used to train the CNN model, and the loss function is as follows: ； in, is the loss function; is the output predicted by the model; is the actual hydrological data; Use the backpropagation algorithm to optimize the weights of the CNN model: ； in, is the updated weight; is the old weight; is the learning rate; is the loss function Weight The partial derivative of Integrate the physical process constraints provided by the SWAT model into the CNN model: ； in, is the final prediction output, is an adjustment function that takes physical constraints into account; Use the Monte Carlo method to estimate the conditional probability distribution of flood forecast results: ； in, The features after fusion The final prediction output probability; is the indicator function; is the number of simulations; Each simulation , use the CNN model to make predictions and get the prediction output; For the Parameter samples for this simulation.

2. The hydrological forecasting method based on deep ensemble learning and physical mechanism according to claim 1, characterized in that: In step S1, preprocessing includes data cleaning and normalization; The normalization formula is as follows: ； in, is the original pixel value or the original time series data value; is the normalized pixel value or the normalized time series data value; It is the minimum value of pixel value or the minimum value of time series data value; It is the maximum value of pixel value or the maximum value of time series data.

3. The hydrological forecasting method based on deep ensemble learning and physical mechanism according to claim 2 is characterized in that: In step S2, the convolutional neural network (CNN) includes a convolutional layer, an activation function, and a pooling layer, and the long short-term memory (LSTM) neural network includes a forget gate, an input gate, a unit state candidate, a unit state update, an output gate, and a hidden state update; Among them, the convolution layer: ； in, Is the position after the convolution operation Output; is the convolution kernel; is the input remote sensing image; Represents the convolution operation; The input image is The pixel value at ; Is the position of the convolution kernel relative to the output feature map Offset The value at 、 Is the position coordinate on the output feature map; 、 is the position coordinate on the convolution kernel; The activation function is the ReLU activation function; The forget gate, input gate, cell state candidate, cell state update, output gate, and hidden state update are as follows: ； ； ； ； ； ； in, is the output of the forget gate; is the sigmoid function; is the weight matrix of the forget gate; represents the vector formed by concatenating the hidden state of the previous time step and the input of the current time step; is the hidden state of the previous time step; is the input of the current time step; is the bias term of the forget gate; is the output of the input gate; is the weight matrix of the input gate; is the bias term of the input gate; Candidate for unit status; is the weight matrix when calculating the unit state candidate; is the bias term of the cell state candidate; is the updated cell state; is the cell state at the previous time step; represents element-wise multiplication; is the output of the output gate; is the weight matrix of the output gate; is the bias term of the output gate; is the updated hidden state.

4. The hydrological forecasting method based on deep ensemble learning and physical mechanism according to claim 3 is characterized in that: In step S5, multiple deep learning models are constructed as basic learners, and the basic learners are trained using the training set data. The specific operations are as follows: Construct multiple base learners and record them as ; Train the base learner: ； in, It is a basic learner The optimal parameters of is the input data; is the true label; is the training set data; Is to use parameters Basic learner For input data predictions.

Images