Basin hydrological multivariable prediction method based on meta-learning TFT architecture

Abstract

The invention discloses a drainage basin hydrological multivariable prediction method based on a meta-learning TFT architecture, and relates to the technical field of hydrological prediction, and the method comprises the steps: collecting multivariable water quality data, external driving force data and static drainage basin attribute data of a plurality of hydrological observation stations; splicing and processing missing values, and dividing the missing values into a training set, a test set and a verification set which are arranged in time as input; the capture of the model on the long-term dependency relationship is enhanced by enhancing the gating circulation unit; capturing a multi-level dependency relationship in the time sequence by using a multi-scale image attention network; and integrating the meta-learning framework, fusing a plurality of subtasks, and outputting a predicted value. Model fusion enhanced meta learning and improved Temporal Fusion Transformer are combined, a multi-scale map attention network and a dynamic model fusion mechanism are fused, the defects of a hydrological prediction model in the aspects of capturing a long-term dependency relationship and generalization ability in a complex watershed environment are overcome, and the precision and reliability of multivariable hydrological prediction are improved.

Description

Technical Field

[0001] The present invention relates to the field of water quality analysis and monitoring, and in particular to a watershed hydrological multivariate prediction method based on a meta-learning TFT architecture. Background Art

[0002] At present, with the increasing global climate change and human activities, the hydrological process in the basin shows a high degree of complexity and nonlinear characteristics. Accurate hydrological prediction is of vital importance to flood prevention and disaster reduction, water resources management, ecological protection and urban planning. However, traditional hydrological prediction methods mainly rely on physical models and statistical models, which face many challenges in dealing with complex basin environments and multivariate data. Although physical models can describe the hydrological process and water cycle of the basin in detail, they usually require a large amount of basin information and parameters, which increases the complexity of model construction and places high demands on data quality. In addition, physical models have low computational efficiency and insufficient adaptability when dealing with nonlinear and dynamically changing hydrological processes. In contrast, although statistical models such as regression analysis and time series models have advantages in computational efficiency, their linear assumptions on data limit their application effect in complex nonlinear systems, making it difficult to fully capture the complex relationship between hydrological variables, resulting in limited prediction accuracy.

[0003] With the improvement of big data and computing power, machine learning and deep learning technologies have gradually been introduced into the field of hydrological prediction. These methods have improved the accuracy and efficiency of hydrological prediction to a certain extent by virtue of their powerful data fitting and pattern recognition capabilities. Specifically, deep learning models such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs) perform well in processing time series data and can effectively capture the temporal dependencies in hydrological data. However, existing machine learning and deep learning methods still have several shortcomings when applied to hydrological multivariate prediction. First, traditional deep learning models perform poorly in processing long-term dependencies and multi-scale spatial dependencies, making it difficult to fully capture the complex dynamic characteristics of the basin hydrological process and achieve accurate long-term predictions. Second, these models have poor generalization capabilities when faced with the diverse characteristics of different basins, and are difficult to adapt to the complex and changing hydrological environment. In addition, the existing methods have limited capabilities in multi-source data fusion and missing value processing, which further restricts the improvement of prediction performance.

[0004] To address the above challenges, researchers have begun to explore methods that combine advanced machine learning techniques such as meta-learning, graph neural networks (GNNs), and attention mechanisms to improve the adaptability and performance of hydrological prediction models. Meta-learning improves the model's ability to quickly adapt to new tasks by training on multiple related tasks; graph neural networks can effectively capture the spatial correlation between observation sites within the basin; and attention mechanisms enhance the model's ability to focus on important features. However, the integration of these technologies still faces problems such as high integration complexity, high consumption of computing resources, and insufficient model stability in practical applications, and further research and optimization are urgently needed. Summary of the invention

[0005] Based on the above-mentioned shortcomings of the prior art, the purpose of the present invention is to provide a basin hydrological multivariate prediction method based on a meta-learning TFT architecture to solve the above-mentioned technical problems.

[0006] To achieve the above object, the present invention provides the following technical solution: a basin hydrological multivariate prediction method based on a meta-learning TFT architecture, comprising:

[0007] S1: Obtain multivariate water quality data, external driving force data, and static watershed attribute data from multiple hydrological observation stations, and construct an observation dataset in the form of a time-station matrix;

[0008] S2: The observation data set is divided into a training set, a test set and a validation set after data preprocessing as input;

[0009] S3: Construct a watershed hydrological multivariate prediction model based on the meta-learning TFT architecture. The modeling steps of the watershed hydrological multivariate prediction model include: inputting the extracted hydrological multivariate weighted features into the graph structure gated residual network; using the reinforced gated recurrent unit to capture the long-term sequence dependency; and passing its output to the multi-scale graph attention network for fusion processing;

[0010] S4: Dynamically weighted fusion of the output with the outputs of multiple subtasks through a dynamic model fusion mechanism;

[0011] S5: Use the model fusion enhanced meta-learning framework to optimize the fusion weights and global model parameters, and generate the final prediction results through the output layer.

[0012] The present invention is further configured as follows: in step S1, multivariate water quality data, external driving force data and static watershed attribute data are presented in the form of a time-site matrix with a daily time resolution, and the multivariate water quality data include dissolved oxygen, runoff and water temperature; the external driving force data include precipitation, snow water equivalent, temperature, water vapor pressure and wind speed; the static watershed attribute data include geographical location, terrain characteristics, climate data, hydrological characteristics, land cover and soil properties, the geographical location is expressed in longitude and latitude, and the terrain characteristics include watershed area, altitude, average slope and snow cover percentage; climate data include annual average precipitation, annual maximum and minimum temperatures, and monthly precipitation and temperature averages; hydrological characteristics include base flow index, annual average flow, different percentile flows and monthly average flow; land cover describes the surface type through the proportion of agriculture, bare land, forest, glacier, lake and urban land; soil properties cover soil depth, sand, silt, clay ratio and organic carbon content, reflecting the soil's ability to penetrate, store and lose water.

[0013] The present invention is further configured that the data preprocessing process in step S2 includes:

[0014] The overall dataset is grouped according to different observation sites;

[0015] Based on the site grouping results, the data were further divided into training set, validation set and test set in a ratio of 7:2:1.

[0016] The present invention is further configured that the graph structure gated residual network in step S3 specifically includes:

[0017] Using graph neural networks to capture the topological relationship between features, the input hydrological multivariate sequence is a time-station matrix In the form of represents the time step, Represents the feature dimension, input feature matrix With the adjacency matrix Combine to form the adjacency relationship of features, and update the features hierarchically through the graph neural network. Layer, calculated by the following update formula:

[0018]

[0019] in, is the activation function, is a trainable weight matrix, which obtains high-level feature representation through iterative calculation ;

[0020] The information flow is dynamically adjusted through the gating mechanism, and the residual connection is used to enhance the feature expression:

[0021]

[0022] in, and is the trainable parameter of the gating layer, is the sigmoid activation function, Represents element-wise multiplication.

[0023] The present invention is further configured that the calculation process of the enhanced gated recurrent unit in step S3 includes:

[0024]

[0025] in, represents the sigmoid function, represents the hyperbolic tangent function, represents element-wise multiplication, , , , and , , is the trainable weight matrix, , , is the bias vector, Indicates the current time step The input data, is the hidden state of the previous moment, indicating the feature representation of the previous time step, is the hidden state at the current moment, indicating the feature representation after being updated at the current time step, is a candidate state, which represents the calculation result based on the current input data and the state at the previous moment. To reset the gate, control the influence of the previous hidden state on the candidate state calculation, To update the gate, determine whether the hidden state at the current moment should be from the candidate state Or is it hidden from the last moment? renew, is the hidden state after residual connection, indicating the current hidden state Combining input data Afterwards, the influence of input information is increased and the ability of feature expression is enhanced.

[0026] The present invention is further configured that the specific implementation of the multi-scale graph attention network in step S3 includes:

[0027] Constructing multiple adjacency matrices at different scales ;

[0028] The graph attention mechanism is used to extract features from the adjacency matrix of each scale. The specific calculation formula is as follows:

[0029]

[0030] in, , , , are query, key, and value matrices respectively, , , is the trainable weight matrix, is the dimension of the key vector.

[0031] The present invention is further configured that the specific implementation of the dynamic model fusion mechanism in step S4 includes:

[0032] Setting multiple subtasks , each subtask has parameters ,in, For the Subtasks The relevant task-specific parameters determine the behavior and output of each subtask model. They are the weights and biases of the subtask model, which are used to adjust how the model responds to the input Make predictions;

[0033] Through the converged network According to the input data Dynamically generate fusion weights , and its calculation formula is:

[0034]

[0035] The output of each subtask is weighted Perform weighted summation to obtain the fused output :

[0036]

[0037] in, For subtask On input The dynamic weight under .

[0038] The present invention is further configured that the optimization process of model fusion enhanced meta-learning in step S5 includes:

[0039] In the inner loop, for each task , using the current fusion model parameters Perform a small amount of gradient updates to get task-specific parameters :

[0040]

[0041] In the outer loop, pass the task-specific parameters of all tasks Calculating Meta-loss , and update the global model parameters and Converged Network :

[0042]

[0043]

[0044]

[0045] in, is the inner loop learning rate, is the outer loop learning rate, is the number of tasks, For the The loss function of each task is and Respectively training and validation sets for each task.

[0046] The present invention provides a basin hydrological multivariate prediction method based on a meta-learning TFT architecture, by establishing a light source characteristic model of a glass-based small-pitch display screen according to the current intensity of an input signal and the adjustment parameters of a backlight source, wherein the light source characteristic model is used to obtain the brightness output at the current moment according to the current intensity of the input signal and the adjustment parameters of the backlight source; the brightness output at the current moment is obtained according to the light source characteristic model, and the operating parameters of the glass-based small-pitch display screen at the current moment are obtained; according to the brightness output and the operating parameters at the current moment, a multidimensional adjustment factor is calculated, and the multidimensional adjustment factor includes a brightness difference factor and a temperature influence factor; a threshold value is judged according to the multidimensional adjustment factor, and when the preset threshold value condition is met, brightness compensation and power optimization are performed, and the beneficial effects produced include:

[0047] 1. Improve the ability to capture long-term dependencies: By integrating the enhanced gated recurrent unit and the TFT architecture into the model, the present invention enhances the model's ability to capture long-term dependencies, and can more accurately predict the long-term trends and changes of basin hydrological variables, overcoming the shortcomings of traditional models in processing long sequence data.

[0048] 2. Multi-level spatial dependency modeling: By adopting a multi-scale graph attention network, the present invention can effectively capture the complex spatial dependencies between observation sites in the basin, enhance the model’s ability to understand and apply multi-level spatial features, and thus improve the accuracy and sophistication of multivariate hydrological forecasts.

[0049] 3. Enhance the generalization ability of the model: By integrating the meta-learning framework, the present invention improves the generalization ability of the model in different watershed environments by integrating the learning of multiple subtasks. This method can quickly adapt and maintain high-precision prediction performance in the case of new watersheds or changes in data distribution, solving the problem that existing models perform poorly under diverse watershed conditions.

[0050] 4. Dynamic model fusion mechanism: The introduction of dynamic model fusion mechanism enables the model to automatically adjust the structure and parameters according to the characteristics of different watershed environments, and realize the efficient integration and processing of multi-source data. This mechanism improves the adaptability and stability of the model, significantly improves the accuracy and reliability of hydrological prediction in complex watershed environments, and reduces the need for manual intervention and parameter tuning.

[0051] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0053] Figure 1 A flowchart of a method for basin hydrological multivariate prediction based on a meta-learning TFT architecture is shown as an exemplary embodiment of the present invention;

[0054] Figure 2 The framework of the basin hydrological multivariate prediction and simulation method based on the meta-learning TFT framework of the present invention;

[0055] Figure 3 It is a time series diagram of the simulated and observed dissolved oxygen values ​​in the watershed of the present invention;

[0056] Figure 4 It is a time series diagram of the simulated value and the observed value of the watershed runoff volume of the present invention;

[0057] Figure 5 It is a time series diagram of the simulated value and observed value of the water temperature in the basin of the present invention. DETAILED DESCRIPTION

[0058] The following will describe the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention, not for limiting the scope of protection of the present invention.

[0059] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0060] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.

[0061] A multivariate prediction method for watershed hydrology based on a meta-learning TFT architecture, such as Figure 1 As shown, including:

[0062] S1: Obtain multivariate water quality data, external driving force data, and static watershed attribute data from multiple hydrological observation stations, and construct an observation dataset in the form of a time-station matrix;

[0063] The acquired water quality observation data, external driving force data and static watershed attribute data are organized in the form of a "time-site matrix" and recorded and processed mainly with daily scale accuracy. First, the static watershed attribute data reflect the constant attributes of each hydrological station, including variables such as snow cover percentage, average annual precipitation, average annual maximum temperature, standard deviation of annual minimum temperature, and average temperature of a specific month, describing the long-term climate and geographical characteristics of the watershed, and are used to analyze the inherent heterogeneity of different stations in the multivariate water quality response process. Secondly, the multivariate water quality time series data includes dissolved oxygen concentration, standardized flow and water temperature water quality indicators, which directly reflect the health status and ecological environment characteristics of the water body. Water temperature is an important factor in controlling dissolved oxygen and aquatic ecological processes in water bodies, and directly affects the physical and chemical properties and ecological environment of water bodies. These water quality time series data are the main target variables of the model, and provide information on the trend of water quality changes over time in the form of a dynamic time-site matrix. External driving force data cover meteorological forcing data on a daily scale, including precipitation, snow water equivalent, average temperature, maximum temperature, minimum temperature, water vapor pressure, and horizontal and vertical components of wind speed. These variables constitute the external driving factors of the spatiotemporal dynamic changes in water quality.

[0064] S2: After data preprocessing, it is divided into training set, test set and validation set as the input of the model;

[0065] For all observation data, we first group them by site, and decompose the data set into independent subsets of each site. The data set of each site contains water quality time series, external driving force data, and static watershed attribute data, and the data is represented in a matrix in the time dimension. The site set is , each site The data can be expressed as a "time-site matrix" ,in is the number of time steps, indicating the observation duration of each station in the dataset, The number of features contained in the observed data for each time step, including water quality parameter series, external driving force data, and static watershed properties.

[0066] After completing the data grouping, in order to ensure the continuity of the time series and the effectiveness of the model training, the data of each site is divided into a training set, a validation set, and a test set in chronological order. Suppose the entire data set is , then the ratio of training set, validation set and test set is 7:2:1. Specifically, the training set is the first 70% of the data , the validation set is the middle 20% of the data , the test set is the last 10% of the data ,Right now:

[0067]

[0068] Apply the time window sliding mechanism to the grouped data for serialization processing. By setting a fixed-length time window , multiple sliding window samples are generated on each site dataset. For each time step , the construction length is The window sequence ,in Indicates time The sliding window mechanism allows the model to learn short-term and long-term dependencies in the time series during training.

[0069] After data partitioning, the multivariate water quality data and external driving force data of each site were standardized to ensure comparability between different characteristics. Calculate the mean and standard deviation , then the original data Convert to standardized data By formula:

[0070]

[0071] In order to deal with the missing values ​​in the original observation data, all missing values ​​are uniformly filled with 0 to form a complete matrix All preprocessed data sets are saved as structured CSV files in the “time-site matrix” format, with time as rows and variables as columns.

[0072] S3: Construct a watershed hydrological multivariate prediction model based on the meta-learning TFT architecture. The improvements of the model include: inputting weighted features into a graph-structured gated residual network to capture complex relationships between features; improving the capture of long-term dependencies by strengthening gated recurrent units; and passing its output to a multi-scale graph attention network to obtain multi-level dependencies in time series;

[0073] like Figure 2 The model input consists of water quality parameter series, external driving force data, and static watershed attribute data, in the form of a “time-site matrix”. In the form of represents the time step, Represents the feature dimension. The input data is extracted through features and time-dependent modeling in different modules to output multivariate water quality prediction values, capturing complex spatiotemporal relationships.

[0074] In this architecture, the topological structure relationship between input features is captured through the graph neural network module. Specifically, the input feature matrix With the adjacency matrix Combine to form the adjacency relationship of features, and update the features hierarchically through the graph neural network. Layer, calculated by the following update formula:

[0075]

[0076] in, is the activation function, is a trainable weight matrix, which obtains high-level feature representation through iterative calculation This process exploits the topological information of the graph structure, enabling the model to understand the spatial relationships between features and adapt to the complex dependencies of watershed features.

[0077] In order to further optimize information transmission and capture the dynamic relationship between features, the gating mechanism and residual connection are introduced to increase the flexibility of information flow and the expressiveness of features. The gating mechanism calculates the gating matrix , dynamically adjusts the information flow between layers and is defined as follows:

[0078]

[0079] in, and is the trainable parameter of the gating layer. The updated features are weighted combined to generate the final residual output:

[0080]

[0081] And calculate the feature expression of the final residual connection:

[0082]

[0083] This gated residual connection structure allows features to retain important information during inter-layer transmission and enhances the depth of feature expression.

[0084] In order to capture the long-term dependencies in the time series, a reinforced gated recurrent unit structure is used. GRU first calculates the update gate and reset gate to adjust the impact of the current input on the hidden state update. The specific calculation is as follows:

[0085]

[0086] The candidate hidden states are then adjusted based on the reset gate:

[0087]

[0088] Finally, the update gate is combined to generate the current hidden state:

[0089]

[0090] After introducing the residual connection, the final hidden state is calculated as:

[0091]

[0092] This GRU structure not only has the time-dependent modeling capability of the standard GRU, but also further enhances feature transfer through residual connections, making the model have stronger representation capabilities for long time series.

[0093] The multi-scale graph attention network is used to capture the multi-level dependencies between features at different scales. First, multiple adjacency matrices are constructed. , and use the graph attention mechanism to extract the features of each scale. In the specific calculation, the query, key and value matrices are represented as , and :

[0094]

[0095] in, , , ,and , , is a trainable parameter. The attention mechanism can dynamically assign weights of different features at different scales to capture multi-scale dependencies and enhance the spatiotemporal expression of features.

[0096] Finally, after feature extraction and integration of graph neural networks, reinforced gated recurrent units, and multi-scale graph attention networks, the model outputs water quality prediction results for multiple future time steps in the time series. This multi-layer structure design enables the model to have high performance in spatial relationship modeling, temporal dependency capture, and multi-scale feature extraction, thereby achieving accurate multivariate basin hydrological prediction.

[0097] S4: Dynamically weighted fusion of the output of the task with the outputs of multiple subtasks through the dynamic model fusion mechanism;

[0098] In step S4, the dynamic model fusion mechanism combines the outputs of multiple subtasks through a fusion network to achieve adaptive weighted fusion, thereby enhancing the overall performance and generalization ability of the model. This mechanism mainly includes the following stages.

[0099] First, multiple subtasks are set , where each subtask With independent parameter set , to adapt to specific subtask requirements and characteristics. These subtasks can be different prediction models or feature extraction modules, aiming to analyze the input data from different perspectives. Modeling.

[0100] Then, through the fusion network Dynamically generate fusion weights , that is, each subtask is in the current input The calculation formula for weight generation uses the softmax function to ensure that the weight satisfies the properties of probability distribution. The formula is as follows:

[0101]

[0102] in, Indicates the first The output activation value of each subtask. is a positive value, and the sum of the weights of all subtasks is 1, that is , ensuring that the fused output is adaptable and stable.

[0103] In the last step, the output of each subtask is weighted according to the dynamic Perform weighted summation to generate the final fusion output The calculation formula is:

[0104]

[0105] In this formula, each subtask The output depends on the current input Weight Perform weighted summation to form the final output prediction This dynamic weighting mechanism can flexibly adapt to the changes in different input features, allowing the model to optimize the combined effect of subtasks under different input conditions, thereby significantly improving the accuracy and robustness in hydrological multivariate prediction.

[0106] S5: Use the model fusion enhanced meta-learning framework to optimize the fusion weights and global model parameters, and generate the final prediction results through the output layer;

[0107] In step S5, the optimization process of model fusion enhanced meta-learning is implemented through two layers of inner and outer loops to optimize the global model parameters and fusion network parameters, so that the model has good adaptability and generalization ability in a multi-task environment. This process combines the inner and outer loop strategies of meta-learning, enabling the model to quickly adapt and optimize global performance on different tasks.

[0108] In the inner loop, for each task , the model uses the current fusion model parameters and converged networks Perform a small amount of gradient updates. Specifically, first calculate the task The loss function , and then adjust the model parameters based on the loss function Perform a gradient descent update to obtain task-specific update parameters This process is achieved through the following formula:

[0109]

[0110] in, is the learning rate of the inner loop, For the task The small amount of gradient updates in this step adapt the model to the task requirements on the task-specific dataset and generate task-specific parameters , thus making the model adaptable in different tasks while keeping the overall structure unchanged.

[0111] In the outer loop, the model passes the task-specific parameters of all tasks Calculate the overall meta-loss , and based on the meta-loss the global parameters and fusion network parameters Update. The meta-loss is defined as follows:

[0112]

[0113] in, Indicates the number of tasks, For the task Meta-loss The performance of each task on the validation set is summarized, providing a measure of the overall performance of the model.

[0114] After calculating the meta-loss, the outer loop updates the global model parameters by gradient descent and fusion network parameters , the specific update formula is as follows:

[0115]

[0116]

[0117] in, is the learning rate of the outer loop. This update process simultaneously optimizes the fusion network parameters and global model parameters , enabling the model to share information between different tasks and improving the overall generalization ability and adaptability of the model.

[0118] After optimization and training in step S5, the model has effectively adjusted the global model parameters through the internal and external loop optimization mechanism. and fusion network parameters , in the inference phase, when new observation data When inputting the model, the fusion network It will first generate dynamic fusion weights based on the input . Weight The calculation of is as follows:

[0119]

[0120] These weights dynamically reflect the current input The degree of dependence on each subtask provides a data-driven weighting mechanism for the final prediction.

[0121] Next, the model will each subtask For input data The predicted output By weight Perform weighted summation to obtain the final output after fusion:

[0122]

[0123] Final Output This is the model's prediction result for multivariate water quality under current input conditions. This output integrates the feature extraction advantages and multi-scale information of different subtasks, and has high spatiotemporal adaptability and prediction robustness while ensuring accuracy.

[0124] Figure 3 , Figure 4 and Figure 5 The prediction results of the model on the test set for dissolved oxygen (DO), standardized flow (Qnorm) and water temperature (WT) are shown respectively. In these figures, the blue curve represents the true value and the orange curve represents the predicted value of the model. It can be observed that the model has achieved high fitting accuracy in the prediction of the three indicators.

[0125] exist Figure 3 The model's predicted values ​​closely follow the fluctuations of the true values. The mean absolute error (MAE) of dissolved oxygen is 0.1429, the root mean square error (RMSE) is 0.2517, and the determination coefficient is It reaches 0.8805. This shows that the model has a strong fitting ability in the periodicity and fluctuation characteristics of dissolved oxygen and can better reflect the actual change trend.

[0126] Figure 4 The prediction results of the standardized flow (Qnorm) are shown. Although the fluctuation of flow is more complex, the model can still capture its changing trend well. The prediction accuracy indicators of standardized flow include: MAE is 0.1564, RMSE is 0.2673, and It is 0.8025, indicating that the model can effectively fit a wider range of flow fluctuations, although the error increases slightly due to the influence of severe fluctuations.

[0127] Figure 5 The prediction results of water temperature (WT) are shown in Figure 1. In this figure, the predicted values ​​of the model almost coincide with the true values, especially in the stable and periodic changes of water temperature, which shows extremely high accuracy. The MAE of water temperature is 0.1497, the RMSE is 0.2494, and The value reaches 0.9050, which reflects that the model captures the periodic characteristics of water temperature most accurately. Figure 3 , Figure 4 and Figure 5 The adaptability and predictive ability of the model on different water quality indicators were verified, especially the excellent fitting effect on dissolved oxygen and water temperature, which proved the robustness and high precision of the model in multivariate hydrological prediction tasks.

[0128] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0129] It should be understood that the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. A and B can be singular or plural. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship, but it may also indicate an "and / or" relationship. Please refer to the context for specific understanding.

[0130] In this application, "at least one" means one or more, and "plurality" means two or more. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can mean: a, b, c, ab, ac, bc, or abc, where a, b, c can be single or multiple.

[0131] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0132] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0133] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0134] In the several embodiments provided in the present application, it should be understood that the disclosed system can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0135] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0136] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0137] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks or optical disks.

[0138] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A basin hydrological multivariate prediction method based on a meta-learning TFT architecture, characterized in that: include: S1: Obtain multivariate water quality data, external driving force data, and static watershed attribute data from multiple hydrological observation stations, and construct an observation dataset in the form of a time-station matrix; S2: The observation data set is divided into a training set, a test set and a validation set after data preprocessing as input; S3: Construct a watershed hydrological multivariate prediction model based on the meta-learning TFT architecture. The modeling steps of the watershed hydrological multivariate prediction model include: inputting the extracted hydrological multivariate weighted features into the graph structure gated residual network; using the reinforced gated recurrent unit to capture the long-term sequence dependency; and passing its output to the multi-scale graph attention network for fusion processing; S4: Dynamically weighted fusion of the output with the outputs of multiple subtasks through a dynamic model fusion mechanism; S5: Use the model fusion enhanced meta-learning framework to optimize the fusion weights and global model parameters, and generate the final prediction results through the output layer.

2. A method for basin hydrological multivariate prediction based on meta-learning TFT architecture according to claim 1, characterized in that: In step S1, multivariate water quality data, external driving force data and static watershed attribute data are presented in the form of a time-station matrix with a daily time resolution. Multivariate water quality data include dissolved oxygen, runoff and water temperature; external driving force data include precipitation, snow water equivalent, temperature, water vapor pressure and wind speed; static watershed attribute data include geographical location, terrain characteristics, climate data, hydrological characteristics, land cover and soil properties. The geographical location is expressed in longitude and latitude. The terrain characteristics include watershed area, altitude, average slope and snow cover percentage; climate data include annual average precipitation, annual maximum and minimum temperatures, and monthly precipitation and temperature averages; hydrological characteristics include base flow index, annual average flow, different percentile flows and monthly average flow; land cover describes the surface type through the proportion of agriculture, bare land, forest, glacier, lake and urban land; soil properties cover soil depth, sand, silt, clay ratio and organic carbon content, reflecting the soil's ability to penetrate, store and lose water.

3. The method for basin hydrological multivariate prediction based on meta-learning TFT architecture according to claim 1 is characterized in that: The data preprocessing process in step S2 includes: The overall dataset is grouped according to different observation sites; Based on the site grouping results, the data were further divided into training set, validation set and test set in a ratio of 7:2:

1.

4. The method for basin hydrological multivariate prediction based on meta-learning TFT architecture according to claim 1 is characterized in that: The graph structure gated residual network in step S3 specifically includes: Using graph neural networks to capture the topological relationship between features, the input hydrological multivariate sequence is a time-station matrix In the form of represents the time step, Represents the feature dimension, input feature matrix With the adjacency matrix Combine to form the adjacency relationship of features, and update the features hierarchically through the graph neural network. Layer, calculated by the following update formula: in, is the activation function, is a trainable weight matrix, which obtains high-level feature representation through iterative calculation ; The information flow is dynamically adjusted through the gating mechanism, and the residual connection is used to enhance the feature expression: in, and is the trainable parameter of the gating layer, is the sigmoid activation function, Represents element-wise multiplication.

5. A method for basin hydrological multivariate prediction based on meta-learning TFT architecture according to claim 4, characterized in that: The calculation process of the enhanced gated recurrent unit in step S3 includes: in, represents the sigmoid function, represents the hyperbolic tangent function, represents element-wise multiplication, , , , and , , is the trainable weight matrix, , , is the bias vector, Indicates the current time step The input data, is the hidden state of the previous moment, indicating the feature representation of the previous time step, is the hidden state at the current moment, indicating the feature representation after being updated at the current time step, is a candidate state, which represents the calculation result based on the current input data and the state at the previous moment. To reset the gate, control the influence of the previous hidden state on the candidate state calculation, To update the gate, determine whether the hidden state at the current moment should be from the candidate state Or is it hidden from the last moment? renew, is the hidden state after residual connection, indicating the current hidden state Combining input data Afterwards, the influence of input information is increased and the ability of feature expression is enhanced.

6. A method for basin hydrological multivariate prediction based on meta-learning TFT architecture according to claim 5, characterized in that: The specific implementation of the multi-scale graph attention network in step S3 includes: Constructing multiple adjacency matrices at different scales ; The graph attention mechanism is used to extract features from the adjacency matrix of each scale. The specific calculation formula is as follows: in, , , , are query, key, and value matrices respectively, , , is the trainable weight matrix, is the dimension of the key vector.

7. The method for basin hydrological multivariate prediction based on meta-learning TFT architecture according to claim 1 is characterized in that: The specific implementation of the dynamic model fusion mechanism in step S4 includes: Setting multiple subtasks , each subtask has parameters ,in, For the Subtasks The relevant task-specific parameters determine the behavior and output of each subtask model. They are the weights and biases of the subtask model, which are used to adjust how the model responds to the input Make predictions; Through the converged network According to the input data Dynamically generate fusion weights , and its calculation formula is: The output of each subtask is weighted Perform weighted summation to obtain the fused output : in, For subtask On input The dynamic weight under .

8. The method for basin hydrological multivariate prediction based on meta-learning TFT architecture according to claim 1 is characterized in that: The optimization process of model fusion enhanced meta-learning in step S5 includes: In the inner loop, for each task , using the current fusion model parameters Perform a small amount of gradient updates to get task-specific parameters : In the outer loop, pass the task-specific parameters of all tasks Calculating Meta-loss , and update the global model parameters and Converged Network : in, is the inner loop learning rate, is the outer loop learning rate, is the number of tasks, For the The loss function of each task is and Respectively training and validation sets for each task.

Images