Flow calibration method and device

Abstract

The invention provides a flow calibration method and device, and belongs to the technical field of hydrological forecasting, and the method comprises the following steps: obtaining geographic information, historical real flow data, historical meteorological data and historical hydrological parameter data in a drainage basin; constructing a space-time variable source mixed runoff production model by using geographic information and historical hydrological parameter data; training the time sequence model by using the historical real flow data, the historical flow prediction data and the historical hydrological parameter data; inputting the target hydrological parameter data and the target meteorological data into a space-time variable source mixed runoff production model to obtain initial flow prediction data; and inputting the initial flow prediction data and the target hydrological parameter data into the time sequence model to obtain calibrated flow prediction data. According to the flow calibration method provided by the invention, the space-time variable source mixed flow generation model is utilized to perform flow preliminary prediction, and then the time sequence model is utilized to calibrate the flow preliminary prediction result, so that the flow prediction result can be effectively improved.

Description

Technical Field

[0001] The invention relates to the technical field of hydrological forecasting, and in particular to a flow calibration method and device. Background Art

[0002] Accurate prediction of hydrological flow is one of the core issues in water resources management, flood warning and environmental protection. In practical applications, flow prediction usually relies on hydrological models, which predict future flow changes by modeling the hydrological processes of the basin. However, due to the complexity of hydrological processes and the heterogeneity of basins, traditional hydrological models often face challenges such as parameter uncertainty, incomplete data and oversimplified models. Summary of the invention

[0003] In view of this, the present invention provides a flow calibration method and device, which are used to solve the current problems of inaccurate hydrological flow prediction results and a large amount of required hydrological data.

[0004] In order to solve the above technical problems, in a first aspect, the present invention provides a flow calibration method, the method comprising:

[0005] Obtain the geographical information of the basin, historical real flow data of each monitoring station, historical meteorological data and historical hydrological parameter data;

[0006] Using the geographic information and the historical hydrological parameter data to construct a spatiotemporal variable source mixed runoff model, inputting the historical meteorological data and the historical hydrological parameter data into the spatiotemporal variable source mixed runoff model for training, and using the spatiotemporal variable source mixed runoff model to perform flow prediction to obtain historical flow prediction data;

[0007] Constructing a time series model, and training the time series model using the historical real flow data, the historical flow prediction data, and the historical hydrological parameter data to obtain a trained time series model;

[0008] Input the target hydrological parameter data and target meteorological data into the trained spatiotemporal variable source mixed flow generation model, use the trained spatiotemporal variable source mixed flow generation model to predict the flow, and obtain preliminary flow prediction data;

[0009] The preliminary flow prediction data and the target hydrological parameter data are input into the trained time series model, and the trained time series model is used to perform flow calibration to obtain calibrated flow prediction data.

[0010] Optionally, before the step of constructing a spatiotemporal variable source mixed runoff generation model using the geographic information and the historical hydrological parameter data, the method further includes:

[0011] The acquired geographic information of the basin, historical real flow data of each monitoring station, historical meteorological data and historical hydrological parameter data are preprocessed to remove abnormal data and interpolate and fill missing values.

[0012] Optionally, the time series model includes a Modern TCN model and a linear transformation layer, and the output of the ModernTCN model is connected to the input of the linear transformation layer.

[0013] Optionally, the number of Modern TCN models included in the time series model is at least two, the input and output of each Modern TCN model are connected in sequence, and the output of the last Modern TCN model is connected to the input of the linear transformation layer.

[0014] Optionally, the Modern TCN model includes a deep convolutional layer and two convolutional feed-forward network layers, the output of the deep convolutional layer is connected to the input of one of the convolutional feed-forward network layers, and the output of the convolutional feed-forward network layer connected to the deep convolutional layer is connected to the input of another convolutional feed-forward network layer.

[0015] In a second aspect, the present invention further provides a flow calibration device, the device comprising:

[0016] The acquisition module is used to obtain the geographical information of the basin, the historical real flow data of each monitoring station, the historical meteorological data and the historical hydrological parameter data;

[0017] The first training module is used to construct a spatiotemporal variable source mixed runoff generation model using the geographic information and the historical hydrological parameter data, input the historical meteorological data and the historical hydrological parameter data into the spatiotemporal variable source mixed runoff generation model for training, and use the spatiotemporal variable source mixed runoff generation model to perform flow prediction to obtain historical flow prediction data;

[0018] A second training module is used to construct a time series model, and train the time series model using the historical real flow data, the historical flow prediction data and the historical hydrological parameter data to obtain a trained time series model;

[0019] A prediction module is used to input the target hydrological parameter data and the target meteorological data into the trained spatiotemporal variable source mixed flow generation model, and use the trained spatiotemporal variable source mixed flow generation model to perform flow prediction to obtain preliminary flow prediction data;

[0020] The calibration module is used to input the preliminary flow prediction data and the target hydrological parameter data into the trained time series model, perform flow calibration using the trained time series model, and obtain calibrated flow prediction data.

[0021] Optionally, the device further comprises:

[0022] The data preprocessing module is used to preprocess the acquired geographic information within the basin, the historical real flow data of each monitoring station, the historical meteorological data and the historical hydrological parameter data, remove abnormal data, and interpolate and fill in missing values.

[0023] Optionally, the time series model includes a Modern TCN model and a linear transformation layer, and the output of the ModernTCN model is connected to the input of the linear transformation layer.

[0024] Optionally, the number of Modern TCN models included in the time series model is at least two, the input and output of each Modern TCN model are connected in sequence, and the output of the last Modern TCN model is connected to the input of the linear transformation layer.

[0025] Optionally, the Modern TCN model includes a deep convolutional layer and two convolutional feed-forward network layers, the output of the deep convolutional layer is connected to the input of one of the convolutional feed-forward network layers, and the output of the convolutional feed-forward network layer connected to the deep convolutional layer is connected to the input of another convolutional feed-forward network layer.

[0026] The beneficial effects of the above technical solution of the present invention are as follows:

[0027] In the embodiment of the present invention, a mixed flow generation model with spatiotemporal variable sources is used to make a preliminary flow prediction, and then a time series model is used to obtain the time characteristics of the data to calibrate the preliminary flow prediction results. This can effectively improve the flow prediction results. In addition, the model requires a small amount of data. By learning from historical data, accurate flow prediction results can be obtained, reducing dependence on detailed observation data. This is particularly suitable for river basins where hydrological parameters are scarce. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A schematic diagram of a flow calibration method in Embodiment 1 of the present invention;

[0029] Figure 2 Schematic diagram of the calculation process of the spatiotemporal variable source mixed runoff generation model in the first embodiment of the present invention;

[0030] Figure 3It is a schematic diagram of the overall structure of a flow calibration device corresponding to the flow calibration method in the first embodiment of the present invention;

[0031] Figure 4 Schematic diagram of the architecture of the Modern TCN model in the first embodiment of the present invention;

[0032] Figure 5 It is a schematic diagram of the structure of a flow calibration device in the second embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solution and advantages of the embodiment of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all of the embodiments. Based on the described embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present invention.

[0034] Accurate prediction of hydrological flow is one of the core issues in water resources management, flood warning and environmental protection. In practical applications, flow prediction usually relies on hydrological models, which predict future flow changes by modeling the hydrological processes of the basin. However, due to the complexity of hydrological processes and the heterogeneity of basins, traditional hydrological models often face challenges such as parameter uncertainty, incomplete data and oversimplified models.

[0035] Calibration of hydrological models is an important step to ensure the accuracy of model predictions. Traditional calibration methods usually rely on optimization algorithms (such as genetic algorithms, particle swarm optimization, etc.) to adjust model parameters to minimize the error between model output and actual observed data. Although these methods can achieve good results in some cases, they are often affected by data quality, computational complexity and model assumptions, and are difficult to cope with variable and complex hydrological environments. Time series models, such as Modern TCN (Modern Temporal Convolutional Network) models, can provide more accurate predictions by capturing the spatiotemporal dependencies in flow data, and then calibrate the prediction results of the initial traditional hydrological model, so that the prediction results are closer to the actual basin monitoring data results.

[0036] In this context, the hybrid method combining traditional hydrological models with time series models can not only improve the prediction accuracy, but also make up for the lack of interpretability of time series models, making the hydrological flow calibration method more comprehensive and reliable, and solving the problem of runoff prediction calibration after constructing a hydrological model in the basin area.

[0037] Therefore, please refer to Figure 1 , Figure 1 A flow chart of a flow calibration method provided in Embodiment 1 of the present invention, the method comprising the following steps:

[0038] Step 11: Obtain the geographic information of the basin, the historical real flow data of each monitoring station, the historical meteorological data and the historical hydrological parameter data.

[0039] In this embodiment, the geographical information in the watershed includes the geographical location, topography, landform, geology, soil, vegetation, etc. of the watershed; the historical meteorological data includes the air pressure, temperature, extreme maximum temperature, extreme minimum temperature, relative humidity, precipitation, evaporation, wind speed, etc. in a certain period in the past; the historical hydrological parameter data includes relevant parameters of the soil and calculation formulas, such as soil moisture, permeability, soil density, evaporation pan coefficient, soil midflow coefficient, soil midflow error and other parameters in a certain period in the past and the corresponding calculation formula parameters.

[0040] Step 12: Use the geographic information and the historical hydrological parameter data to construct a spatiotemporal variable source mixed runoff generation model, input the historical meteorological data and the historical hydrological parameter data into the spatiotemporal variable source mixed runoff generation model for training, and use the spatiotemporal variable source mixed runoff generation model to predict flow to obtain historical flow prediction data.

[0041] In the embodiment of the present invention, the traditional hydrological model selects a distributed hydrological model, specifically a time-space variable source mixed runoff model. The time-space variable source mixed runoff model is a model specifically used to describe and simulate hydrological processes such as groundwater flow and mountain torrents. It combines the characteristics of time series models and spatial models, and takes into account the source-sink process and time-space changes in the groundwater flow system. The basic idea of ​​the model is to divide the groundwater flow system or the study area into several small areas, and the source-sink process in each small area is described by the runoff model, while the time-space changes between different small areas are captured and described by time series models and spatial models.

[0042] In this step, a simulation model of the hydrological process, i.e., a spatiotemporal variable source mixed runoff model, can be constructed by combining the acquired geographic information in the basin with the collected historical hydrological parameter data. Then, the historical meteorological data and the historical hydrological parameter data are input into the spatiotemporal variable source mixed runoff model for training. The trained spatiotemporal variable source mixed runoff model can be used to predict the flow based on the historical meteorological data and the historical hydrological parameter data to obtain the corresponding historical flow prediction data, which are used as the basic data for the subsequent time series model.

[0043] Step 13: construct a time series model, and train the time series model using the historical real flow data, the historical flow prediction data, and the historical hydrological parameter data to obtain a trained time series model.

[0044] Since the data features in the spatiotemporal variable source mixed runoff model contain time features, a time series model can be further constructed to extract the time features. Specifically, a time series model is first constructed, and then the time series model is trained using historical real flow data, historical flow prediction data, and historical hydrological parameter data to obtain a trained time series model. The trained time series model can be used to calibrate the preliminary flow prediction results output by the spatiotemporal variable source mixed runoff model.

[0045] Optionally, the time series model uses a Modern TCN model, which receives historical flow prediction data and historical hydrological parameter data, improves data processing efficiency through one-dimensional convolution operations, and can effectively capture long-distance time dependencies through causal convolution and extended convolution to avoid gradient vanishing or gradient exploding.

[0046] In this embodiment, whether the time series model is trained can be determined based on the difference between the traffic prediction data output by the time series model and the corresponding historical real traffic data.

[0047] Step 14: Input the target hydrological parameter data and target meteorological data into the trained spatiotemporal variable source mixed runoff model, use the trained spatiotemporal variable source mixed runoff model to predict the flow, and obtain preliminary flow prediction data.

[0048] In this step, after completing the construction and training of the spatiotemporal variable source mixed runoff model and the time series model used to calibrate the output results of the spatiotemporal variable source mixed runoff model, the acquired target hydrological parameter data and target meteorological data can be input into the trained spatiotemporal variable source mixed runoff model. The trained spatiotemporal variable source mixed runoff model can perform flow prediction based on the input data and then output preliminary flow prediction data.

[0049] Step 15: Input the preliminary flow prediction data and the target hydrological parameter data into the trained time series model, and use the trained time series model to perform flow calibration to obtain calibrated flow prediction data.

[0050] In this embodiment, further, it is necessary to input the preliminary flow prediction data and target hydrological parameter data output by the spatiotemporal variable source mixed flow generation model into the trained time series model, and use the trained time series model to calibrate the flow, and finally obtain calibrated and more accurate flow prediction data.

[0051] Optionally, the flow prediction data in this embodiment may include the river flow and the moisture content of the soil in the river at the corresponding moment.

[0052] The flow calibration method provided in the embodiment of the present invention uses a spatiotemporal variable source mixed flow generation model to make a preliminary flow prediction, and then uses a time series model to obtain the time characteristics of the data to calibrate the preliminary flow prediction results, which can effectively improve the flow prediction results. In addition, the model requires a small amount of data. By learning from historical data, accurate flow prediction results can be obtained, reducing dependence on detailed observation data. It is particularly suitable for river basins where hydrological parameters are scarce.

[0053] The following example illustrates the above flow calibration method.

[0054] In one optional specific implementation manner, before the step of using the geographic information and the historical hydrological parameter data to construct a spatiotemporal variable source mixed runoff generation model, the method further includes:

[0055] The acquired geographic information of the basin, historical real flow data of each monitoring station, historical meteorological data and historical hydrological parameter data are preprocessed to remove abnormal data and interpolate and fill missing values.

[0056] In this embodiment, after collecting historical real flow data, meteorological data (such as rainfall, evaporation, etc.) and roughly measured hydrological parameter data (such as soil moisture, permeability, etc.) from various monitoring stations in the basin, the collected data needs to be cleaned to remove obviously erroneous or abnormal data points. Among them, for missing values, appropriate interpolation methods (such as linear interpolation, spline interpolation, etc.) are used to fill in to ensure the continuity and integrity of the data. In this way, the prediction accuracy of the model can be further improved.

[0057] In some embodiments, the time series model includes a Modern TCN model and a linear transformation layer, and the output of the Modern TCN model is connected to the input of the linear transformation layer.

[0058] In this embodiment, the Modern TCN model is used to make time series predictions of flow and related hydrological characteristics. The ModernTCN model is a modern transformation of the traditional TCN (Temporal Convolutional Network) by adding a deep convolutional network (DW Conv) and a convolutional feedforward network (Conv FNN) to improve the efficiency of the model in processing time series data and capture long-distance temporal dependencies.

[0059] The linear transformation layer maps the input data to a new space through a linear transformation. This linear transformation is achieved by matrix multiplication and adding a bias. In this way, the linear transformation layer can change the dimension of the data for further processing and output calibrated traffic prediction data.

[0060] In some embodiments, the number of Modern TCN models included in the time series model is at least two, the input and output of each Modern TCN model are connected in sequence, and the output of the last Modern TCN model is connected to the input of the linear transformation layer.

[0061] That is to say, in this embodiment, the number of Modern TCN models used in the time series model is at least two, and the input and output of at least two Modern TCN models are connected in sequence, that is, the input of one Modern TCN model is connected to the output of another Modern TCN model, and the output of the last Modern TCN model is connected to the input of the linear transformation layer. Through multiple Modern TCN models, the time series characteristics of the input data can be effectively extracted, thereby improving the calibration effect of the preliminary flow prediction data input to the spatiotemporal variable source mixed flow generation model.

[0062] In some other embodiments, the Modern TCN model includes a deep convolutional layer and two convolutional feed-forward network layers, the output of the deep convolutional layer is connected to the input of one of the convolutional feed-forward network layers, and the output of the convolutional feed-forward network layer connected to the deep convolutional layer is connected to the input of another convolutional feed-forward network layer.

[0063] Specifically, the depth wise convolution (DW Conv) layer uses a lightweight convolution operation. This network layer can independently learn the temporal dependency of each univariate time series, and uses a large number of convolution cores to increase the effective receptive field, that is, to increase the range of the input area that the current node can actually receive. In other words, the deep convolution network is responsible for learning the temporal information between tags based on each feature, which plays the same role as the self-attention module in the Transformer, and uses a larger convolution kernel, which helps to expand the convolution kernel and improve the effective receptive field of the model, so as to better capture the dependency of the time series.

[0064] The Convolutional Feed-Forward Network (Conv FFN) is a structure that introduces convolution operations into the traditional fully connected feedforward neural network (FFN). It combines the local feature extraction capability of the Convolutional Neural Networks (CNN) and the global feature processing capability of the fully connected layer, and improves the perception of local features by improving the feedforward network in the Transformer structure. Specifically, the Convolutional Feed-Forward Network layer consists of two point convolution convolution networks (PW Conv) and adopts an inverted bottleneck structure. The hidden channel of the Convolutional Feed-Forward Network is wider than the input channel, and it can independently learn the new characteristics of each feature, realizing the separation of visual information mixing and feature information mixing. Each deep convolutional network and convolutional feed-forward network only mixes information in one time or feature dimension. Unlike the traditional convolution that jointly mixes information in two dimensions, this decoupled design can make object tasks easier to learn and reduce computational complexity.

[0065] The following example continues to illustrate the above flow calibration method

[0066] Please refer to Figure 2 , Figure 2 FIG. 1 is a schematic diagram of the calculation process of the spatiotemporal variable source mixed runoff generation model in the first embodiment of the present invention. Figure 2 As shown, the spatiotemporal variable source mixed runoff generation model in this embodiment is based on the nonlinear infiltration calculation method of the vadose zone soil of the one-dimensional infiltration theory, that is, the time-varying infiltration capacity of the surface soil of the basin hydrological response unit is solved by the method of "discrete soil moisture content to calculate the downward movement of the wetting front", the main parameter characteristics of different soil types are analyzed, and a plane, vertical, and time period mixed runoff generation model is established.

[0067] First, meteorological data such as temperature, precipitation, radiation and wind speed are input to provide basic parameter conditions for subsequent hydrological process simulation. The Digital Elevation Model (DEM) in the grid data depicts the topography of the basin, which is crucial for determining the direction of water flow and calculating slope runoff. Data such as land use, soil texture and vegetation cover affect processes such as infiltration, interception and evaporation.

[0068] Grid data is a data representation method that divides geographic space into a series of continuous, equal-sized grid cells and assigns corresponding attribute values ​​or variable values ​​to each grid cell. In the spatiotemporal variable source mixed runoff model, grid data is used to describe and simulate the spatiotemporal variation process in geographic space, such as groundwater flow, rainfall distribution, pollutant diffusion, etc.

[0069] Pre-processing methods such as meteorological difference and precipitation correction can be used to further enhance the accuracy and rationality of the data, so that subsequent hydrological calculations can be based on reliable data.

[0070] Through evaporation, interception, infiltration and soil water movement, the entire process of the hydrological cycle is completely simulated, and finally the excess surface runoff, full runoff and preferential flow, lake and reservoir confluence, and slope confluence are integrated into the river flow to ultimately form the output of total runoff.

[0071] Please refer to Figure 3 and Figure 4 , Figure 3 Schematic diagram of the overall structure of a flow calibration device corresponding to the flow calibration method in Embodiment 1 of the present invention, Figure 4 Schematic diagram of the architecture of the Modern TCN model in the first embodiment of the present invention.

[0072] The input data include basic parameters of the hydrological model and external monitoring parameters. The basic parameters of the hydrological model include the area of ​​distributed hydrological units, radiation coefficient, soil unit parameters, shortwave radiation parameters, evapotranspiration coefficient, environmental interception coefficient, runoff index, etc., which correspond to hydrological parameter data (including soil-related parameters and calculation formula parameters, etc.); external monitoring parameters include rainfall time series (hourly records) at the monitoring site, local temperature records (hourly records), and outlet flow (hourly records) of the upstream river basin on which the hydrological model basin depends, which correspond to meteorological data, etc.

[0073] Before inputting data into the model, the input data can be divided into dynamic vectors and static vectors according to whether the basin characteristics have changed. Dynamic vectors refer to characteristics that change over time, such as maximum temperature, minimum temperature, rainfall, etc., while static vectors refer to basin characteristics that remain fixed for a long time, such as soil type, initial soil moisture content, climate type, groundwater flow, etc. After the initial input of dynamic vectors and static vectors is completed, they constitute the initial input parameter file of the model.

[0074] The hydrological model adopts the spatiotemporal variable source mixed runoff model (referred to as the spatiotemporal variable source model). The calculation logic includes more than ten steps such as meteorological difference and precipitation correction. The calculation evolution is carried out step by step from rainfall to river flow. For details, please refer to the above content and Figure 2 The spatiotemporal variable source mixed flow generation model processes the input data and outputs preliminary flow prediction data.

[0075] The time series model (hereinafter referred to as the time series model) uses the Modern TCN model to make time series predictions of flow and related hydrological characteristics, modernizes the traditional TCN part, adds a deep convolutional network (DW Conv) and a convolutional feedforward network (Conv FNN), improves the efficiency of the model in processing time series data, and captures long-distance time dependencies. In this embodiment, the time series model includes two Modern TCN models.

[0076] Each Modern TCN model includes a deep convolutional layer and two convolutional feedforward network layers, and each convolutional feedforward network layer includes two point convolutional layers. The output of the deep convolutional layer needs to be batch normalized (Batch Normalization, BN). By normalizing the input data of each batch, the data distribution is made more stable, thereby accelerating the convergence speed of the gradient descent method. Gaussian Error Linear Units (GeLU) are introduced between the two point convolutional layers. By introducing error linear units, the model can learn more complex functional relationships.

[0077] The time series model processes the input data, specifically performs flow calibration, and outputs calibrated flow forecast data.

[0078] The output data includes the calibrated river flow and the moisture content of the soil in the river at the corresponding moment.

[0079] In summary, the embodiment of the present invention provides a flow calibration method based on a time series model (Modern TCN model) and a traditional hydrological model (temporal and spatial variable source mixed flow generation model) to address the shortcomings of traditional hydrological models in dealing with complex hydrological processes, data scarcity, and over-simplification of models. Traditional hydrological flow calibration methods usually rely on detailed physical models and a large amount of field observation data. In some basins with insufficient or complex hydrological parameter data, traditional methods alone may not be able to effectively perform accurate parameter measurements. The present invention combines the time series modeling capabilities in deep learning and utilizes the spatiotemporal correlation in flow data to overcome the limitations of traditional models. The method can learn the laws of hydrological processes through the historical predicted flow data of traditional hydrological models and the real river monitoring flow data. Subsequently, only the flow data of the traditional model and the roughly measured hydrological parameters need to be input into the model to obtain accurate flow prediction results; at the same time, the prediction benchmark is determined by the traditional hydrological model, and the flow is calibrated by the time series model, which ensures the interpretability and accuracy of the current model, thereby producing a universal flow calibration model that can achieve accurate flow calibration and prediction by adjusting model parameters and adapting to different basin characteristics.

[0080] The present invention can provide more accurate hydrological flow predictions to help managers plan the allocation and use of water resources in advance. For example, in agricultural irrigation, the irrigation time and water volume can be reasonably arranged according to the accurate flow prediction to improve the efficiency of water resource utilization; it can also accurately predict the flood flow in advance, so that relevant departments have more time to take flood prevention measures, such as strengthening dams, evacuating people, etc., to minimize the disaster losses caused by floods. In terms of academic research, the present invention provides new tools and methods for studying the hydrological cycle process. Scholars can use this model to simulate hydrological changes under different climate and rainfall conditions, and gain a deeper understanding of the impact mechanism of climate change on the hydrological cycle.

[0081] See also Figure 5 , Figure 5 : is a schematic diagram of the structure of a flow calibration device provided in Embodiment 2 of the present invention, the device 50 comprises:

[0082] An acquisition module 51 is used to acquire geographic information within the basin, historical real flow data of each monitoring station, historical meteorological data and historical hydrological parameter data;

[0083] The first training module 52 is used to construct a spatiotemporal variable source mixed runoff model using the geographic information and the historical hydrological parameter data, input the historical meteorological data and the historical hydrological parameter data into the spatiotemporal variable source mixed runoff model for training, and use the spatiotemporal variable source mixed runoff model to perform flow prediction to obtain historical flow prediction data;

[0084] The second training module 53 is used to construct a time series model, and train the time series model using the historical real flow data, the historical flow prediction data and the historical hydrological parameter data to obtain a trained time series model;

[0085] The prediction module 54 is used to input the target hydrological parameter data and the target meteorological data into the trained spatiotemporal variable source mixed flow generation model, and use the trained spatiotemporal variable source mixed flow generation model to perform flow prediction to obtain preliminary flow prediction data;

[0086] The calibration module 55 is used to input the preliminary flow prediction data and the target hydrological parameter data into the trained time series model, perform flow calibration using the trained time series model, and obtain calibrated flow prediction data.

[0087] Optionally, the device further comprises:

[0088] The data preprocessing module is used to preprocess the acquired geographic information within the basin, the historical real flow data of each monitoring station, the historical meteorological data and the historical hydrological parameter data, remove abnormal data, and interpolate and fill in missing values.

[0089] Optionally, the time series model includes a Modern TCN model and a linear transformation layer, and the output of the ModernTCN model is connected to the input of the linear transformation layer.

[0090] Optionally, the number of Modern TCN models included in the time series model is at least two, the input and output of each Modern TCN model are connected in sequence, and the output of the last Modern TCN model is connected to the input of the linear transformation layer.

[0091] Optionally, the Modern TCN model includes a deep convolutional layer and two convolutional feed-forward network layers, the output of the deep convolutional layer is connected to the input of one of the convolutional feed-forward network layers, and the output of the convolutional feed-forward network layer connected to the deep convolutional layer is connected to the input of another convolutional feed-forward network layer.

[0092] In the embodiment of the present invention, a mixed flow generation model with spatiotemporal variable sources is used to make a preliminary flow prediction, and then a time series model is used to obtain the time characteristics of the data to calibrate the preliminary flow prediction results. This can effectively improve the flow prediction results. In addition, the model requires a small amount of data. By learning from historical data, accurate flow prediction results can be obtained, reducing dependence on detailed observation data. This is particularly suitable for river basins where hydrological parameters are scarce.

[0093] The embodiment of the present invention is a product embodiment corresponding to the above-mentioned method embodiment 1, so it will not be described in detail here. Please refer to the above-mentioned embodiment 1 for details.

[0094] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A flow calibration method, characterized in that: The method comprises the following steps: Obtain the geographical information of the basin, historical real flow data of each monitoring station, historical meteorological data and historical hydrological parameter data; Using the geographic information and the historical hydrological parameter data to construct a spatiotemporal variable source mixed runoff model, inputting the historical meteorological data and the historical hydrological parameter data into the spatiotemporal variable source mixed runoff model for training, and using the spatiotemporal variable source mixed runoff model to perform flow prediction to obtain historical flow prediction data; Constructing a time series model, and training the time series model using the historical real flow data, the historical flow prediction data, and the historical hydrological parameter data to obtain a trained time series model; Input the target hydrological parameter data and target meteorological data into the trained spatiotemporal variable source mixed flow generation model, use the trained spatiotemporal variable source mixed flow generation model to predict the flow, and obtain preliminary flow prediction data; The preliminary flow prediction data and the target hydrological parameter data are input into the trained time series model, and the trained time series model is used to perform flow calibration to obtain calibrated flow prediction data.

2. The method according to claim 1, characterized in that Before the step of using the geographic information and the historical hydrological parameter data to construct a spatiotemporal variable source mixed runoff generation model, the method further includes: The acquired geographic information of the basin, historical real flow data of each monitoring station, historical meteorological data and historical hydrological parameter data are preprocessed to remove abnormal data and interpolate and fill missing values.

3. The method according to claim 1, characterized in that The time series model includes a Modern TCN model and a linear transformation layer, and the output of the Modern TCN model is connected to the input of the linear transformation layer.

4. The method according to claim 3, characterized in that The number of Modern TCN models included in the time series model is at least two, the input and output of each Modern TCN model are connected in sequence, and the output of the last Modern TCN model is connected to the input of the linear transformation layer.

5. The method according to claim 3, characterized in that The Modern TCN model includes a deep convolutional layer and two convolutional feed-forward network layers, the output of the deep convolutional layer is connected to the input of one of the convolutional feed-forward network layers, and the output of the convolutional feed-forward network layer connected to the deep convolutional layer is connected to the input of another convolutional feed-forward network layer.

6. A flow calibration device, characterized in that: The device comprises: The acquisition module is used to obtain the geographical information of the basin, the historical real flow data of each monitoring station, the historical meteorological data and the historical hydrological parameter data; The first training module is used to construct a spatiotemporal variable source mixed runoff generation model using the geographic information and the historical hydrological parameter data, input the historical meteorological data and the historical hydrological parameter data into the spatiotemporal variable source mixed runoff generation model for training, and use the spatiotemporal variable source mixed runoff generation model to perform flow prediction to obtain historical flow prediction data; A second training module is used to construct a time series model, and train the time series model using the historical real flow data, the historical flow prediction data and the historical hydrological parameter data to obtain a trained time series model; A prediction module is used to input the target hydrological parameter data and the target meteorological data into the trained spatiotemporal variable source mixed flow generation model, and use the trained spatiotemporal variable source mixed flow generation model to perform flow prediction to obtain preliminary flow prediction data; The calibration module is used to input the preliminary flow prediction data and the target hydrological parameter data into the trained time series model, perform flow calibration using the trained time series model, and obtain calibrated flow prediction data.

7. The device according to claim 6, characterized in that The device also includes: The data preprocessing module is used to preprocess the acquired geographic information within the basin, the historical real flow data of each monitoring station, the historical meteorological data and the historical hydrological parameter data, remove abnormal data, and interpolate and fill in missing values.

8. The device according to claim 6, characterized in that The time series model includes a Modern TCN model and a linear transformation layer, and the output of the Modern TCN model is connected to the input of the linear transformation layer.

9. The device according to claim 8, characterized in that The number of Modern TCN models included in the time series model is at least two, the input and output of each Modern TCN model are connected in sequence, and the output of the last Modern TCN model is connected to the input of the linear transformation layer.

10. The device according to claim 8, characterized in that The Modern TCN model includes a deep convolutional layer and two convolutional feed-forward network layers, the output of the deep convolutional layer is connected to the input of one of the convolutional feed-forward network layers, and the output of the convolutional feed-forward network layer connected to the deep convolutional layer is connected to the input of another convolutional feed-forward network layer.

Images