Underground water level prediction method based on combination of multi-source data and deep learning model

Abstract

The invention discloses an underground water level prediction method based on combination of multi-source data and a deep learning model. The method comprises the following steps: acquiring underground water level monitoring data and multi-source characteristic variable data of a research area; performing spatial resolution standardization on the underground water level monitoring data and the multi-source characteristic variable data to obtain a standard data set with unified spatial resolution; the standard data set is subjected to normalization processing, normalized data is utilized to train a deep learning model, and the deep learning model comprises a long and short term memory network model, a graph convolution network model and a Transform model; and predicting the underground water level at the future moment by using the trained prediction model. The method solves the problem that the traditional groundwater level prediction method does not consider the difference of different data sets in the aspects of spatial resolution, time resolution, observation error and the like, and the precision and robustness of the model are difficult to improve.

Description

Technical Field

[0001] This invention belongs to the field of water conservancy engineering technology, and in particular relates to a groundwater level prediction method based on a combination of multi-source data and deep learning models. Background Technology

[0002] Groundwater is a vital natural resource that supports the sustainable development of human society and maintains ecological stability. Globally, groundwater plays an irreplaceable role in ensuring water security, maintaining ecosystem stability, and improving human well-being.

[0003] Current groundwater level prediction methods often directly use multi-source datasets as feature variables to build models. However, these multi-source datasets often come from different sources, resulting in significant heterogeneity in spatial resolution, temporal resolution, and observation error distribution. Furthermore, existing prediction techniques mostly use single-structure prediction models for fitting and prediction.

[0004] However, the existing technologies still have the following technical problems: traditional groundwater level prediction methods do not take into account the differences in spatial resolution, temporal resolution and observation error of different datasets, making it difficult to improve the accuracy and robustness of the model. Furthermore, the existing prediction models are relatively simple, which is not conducive to improving the applicability of the model under different hydrogeological conditions and management scenarios. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a groundwater level prediction method based on a combination of multi-source data and a deep learning model. This method solves the problems of traditional groundwater level prediction methods failing to consider the differences in spatial resolution, temporal resolution, and observation errors among different datasets, making it difficult to improve the accuracy and robustness of the model. Furthermore, existing prediction models are relatively simple, which is not conducive to improving the applicability of the model under different hydrogeological conditions and management scenarios.

[0006] To achieve the above objectives, this invention provides a groundwater level prediction method based on a combination of multi-source data and a deep learning model, comprising: Acquire groundwater level monitoring data and multi-source characteristic variable data for the study area; Spatial resolution standardization was performed on the groundwater level monitoring data and multi-source characteristic variable data to obtain a standard dataset with uniform spatial resolution. The standard dataset is normalized, and a deep learning model is trained using the normalized data. The deep learning model includes a long short-term memory network model, a graph convolutional network model, and a Transformer model. The trained prediction model is used to predict the groundwater level at future times.

[0007] Optionally, the multi-source feature variable data includes: meteorological data, remote sensing data, and human activity data; The meteorological data includes: precipitation, temperature, and potential evapotranspiration; The remote sensing data includes: normalized vegetation index, surface temperature, surface elevation, and soil moisture content; The human activity data includes: land use type and nighttime light index.

[0008] Optionally, the groundwater level monitoring data and multi-source characteristic variable data are spatially standardized to obtain a standard dataset with uniform spatial resolution, including: The groundwater level monitoring data is interpolated using ordinary kriging to obtain raster data with a preset spatial resolution, thus acquiring the first standard data. For multi-source feature variable data whose spatial resolution is not the preset spatial resolution, bilinear interpolation and majority class resampling are used to resample them to the preset spatial resolution to obtain second standard data. Based on the first and second standard data, a standard dataset with unified spatial resolution is obtained.

[0009] Optionally, it can be resampled to a preset spatial resolution using bilinear interpolation and majority class resampling, including: Bilinear interpolation is used to sample data from the meteorological data, remote sensing data, and nighttime light index that are not at the preset resolution to the preset spatial resolution. The majority class resampling method is used to sample data from the land use type that are not at the preset resolution to the preset spatial resolution.

[0010] Optionally, normalizing the standard dataset includes: ; in, For the normalized data, For initial data, and These are the maximum and minimum values ​​in the dataset, respectively.

[0011] Optionally, the long short-term memory network model includes: a bottom layer, an intermediate layer, a top layer, and a fully connected output layer; The underlying layer is used to extract the primary hidden state at the time step; The intermediate layer is used to randomly discard neurons in the primary hidden state with a preset probability; The top layer is used to extract high-level temporal features based on the discarded hidden state; The fully connected output layer is used to map the high-level temporal features into single-step water level prediction values.

[0012] Optionally, the graph convolutional network model includes: a multi-factor adjacency matrix construction layer, a two-layer graph convolutional layer, and a fully connected output layer; The multi-factor adjacency matrix construction layer is used to obtain a comprehensive adjacency matrix by integrating three dimensions: geographical distance, elevation difference, and trend correlation. The dual-layer graph convolutional layer is used to perform first-order and second-order neighborhood information aggregation on the comprehensive adjacency matrix. The fully connected output layer is used to map the aggregated node features into single-step water level prediction values.

[0013] Optionally, the Transformer model includes: an embedding layer, a position encoding layer, a Transformer encoder, and an output layer; The embedding layer is used to linearly project multi-dimensional, multi-source features onto an embedding vector of uniform dimension. The location encoding layer is used to add learnable temporal location information to the embedding vector; The Transformer encoder is used to capture global temporal dependencies through a self-attention mechanism; The output layer is used to compress the encoder output into a single-step water level prediction value.

[0014] Optionally, obtaining a trained prediction model includes: The normalized data is input into the Long Short-Term Memory Network Model, the Graph Convolutional Network Model, and the Transformer Model. The Adam optimizer and the mean squared error loss function are used to iteratively update each model to obtain the trained Long Short-Term Memory Network Model, the Graph Convolutional Network Model, and the Transformer Model. Based on the trained Long Short-Term Memory (LSTM) network model, Graph Convolutional Network (GCN) model, and Transformer model, and combined with the multi-source feature variable dataset, we obtain the LSM network predicted water level, the Graph Convolutional Network predicted water level, and the Transformer model predicted water level, which correspond one-to-one with the spatial grid. Based on the water level predictions of the Long Short-Term Memory Network, Graph Convolutional Network, and Transformer Model, the mean absolute error of the Long Short-Term Memory Network, the mean absolute error of the Graph Convolutional Network, and the mean absolute error of the Transformer Model at each spatial grid are calculated. The model with the smallest mean absolute error at each spatial grid is marked as the optimal model for that grid. The optimal models of all grids are then spliced ​​together according to their spatial positions to form an optimal model combination covering the entire study area. This optimal model combination is then used as the trained prediction model.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: 1. In the data processing stage of this invention, groundwater level data is interpolated to a dataset with a spatial resolution of 0.1°×0.1° using the ordinary Kriging interpolation method. For multi-source datasets, bilinear interpolation is used to unify data with different spatial resolutions, while land use type data is processed using the majority class resampling method. This effectively solves the problem of inconsistent spatial resolutions in multi-source data and improves the accuracy and robustness of prediction results.

[0016] 2. This invention constructs three deep learning models: LSTM, GCN, and Transformer. It adopts the Adam optimizer, selects the mean squared error as the loss function of the model, introduces the EarlyStopping callback mechanism to prevent the model from overfitting, and finds the optimal hyperparameter configuration through Bayesian optimization.

[0017] 3. This invention evaluates the prediction performance of different deep learning models based on mean absolute error, and selects the model with the smallest prediction error in different sub-regions to form the optimal combination of groundwater level prediction models for the entire region. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a groundwater level prediction method based on a combination of multi-source data and a deep learning model, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of groundwater level data interpolated to a spatial resolution of 0.1°×0.1° according to an embodiment of the present invention; Figure 3 This is a schematic diagram of land use type data resampled to 0.1°×0.1° spatial resolution data according to an embodiment of the present invention; Figure 4 This is a schematic diagram of NDVI data interpolation with a spatial resolution of 0.1° × 0.1° according to an embodiment of the present invention; Figure 5 This is a schematic diagram of soil moisture content data interpolated to a spatial resolution of 0.1°×0.1° according to an embodiment of the present invention; Figure 6 This is a diagram showing the mean absolute error distribution of the LSTM model in an embodiment of the present invention. Figure 7 This is a diagram showing the mean absolute error distribution of the GCN model in an embodiment of the present invention; Figure 8 This is a diagram showing the mean absolute error distribution of the Transformer model in an embodiment of the present invention; Figure 9 This is a spatial distribution map of the optimal prediction model in an embodiment of the present invention; Figure 10 This is a spatial distribution diagram of the average absolute error of the optimal model combination in this embodiment of the invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0021] This embodiment proposes a groundwater level prediction method based on a combination of multi-source data and a deep learning model, such as... Figure 1 As shown, the specific steps include: Acquire groundwater level monitoring data and multi-source characteristic variable data for the study area; Spatial resolution standardization was performed on the groundwater level monitoring data and multi-source characteristic variable data to obtain a standard dataset with uniform spatial resolution. The standard dataset is normalized, and a deep learning model is trained using the normalized data. The deep learning model includes a long short-term memory network model, a graph convolutional network model, and a Transformer model. The trained prediction model is used to predict the groundwater level at future times.

[0022] Specifically, S1, acquire groundwater level monitoring data in the study area, and acquire multi-source data in the study area, including meteorological data, remote sensing data, and human activity data; S2. Standardize the data with different spatial resolutions and organize them into a standard dataset with a unified spatial resolution. S3. Normalize the groundwater level monitoring data and multi-source datasets, and divide the normalized data into training set, validation set and test set; S4. Construct three deep learning models: Long Short-Term Memory Network (LSTM), Graph Convolutional Network (GCN), and Transformer model. Train the models by inputting them into the training set and test set, and evaluate the model prediction performance. S5. Based on the model's prediction performance, select the optimal model for each sub-region to form the optimal model combination.

[0023] More specifically, the study first acquired groundwater level data from monitoring stations within the study area. Then, it collected meteorological, remote sensing, and human activity data to form a multi-source dataset. The meteorological data included precipitation, temperature, and potential evapotranspiration; the remote sensing data included normalized difference vegetation index (NDVI), surface temperature, surface elevation, and soil moisture content; and the human activity data included land use type and nighttime light index. For groundwater level data, ordinary kriging interpolation was used to interpolate the independent station monitoring data into 0.1°×0.1° raster data. For multi-source datasets with spatial resolutions other than 0.1°×0.1°, a dual-resolution interpolation method was employed. Linear interpolation was used to interpolate the data into a dataset with a uniform spatial resolution. Land use type data, which needed to maintain its integer type, was processed using majority class resampling. Min-Max normalization was used to map the data to the [0,1] interval, and the dataset was divided into training, validation, and test sets in a ratio of 0.7:0.1:0.2. Then, three deep learning models, LSTM, GCN, and Transformer, were constructed and used as inputs to train and optimize the models for groundwater level prediction. Finally, the optimal model for each sub-region was selected by mean absolute error, and the optimal combination of groundwater level prediction models covering the study area was constructed.

[0024] Furthermore, the multi-source feature variable data includes: meteorological data, remote sensing data, and human activity data; The meteorological data includes: precipitation, temperature, and potential evapotranspiration; The remote sensing data includes: normalized vegetation index, surface temperature, surface elevation, and soil moisture content; The human activity data includes: land use type and nighttime light index.

[0025] Furthermore, the groundwater level monitoring data and multi-source characteristic variable data are spatially standardized to obtain a standard dataset with unified spatial resolution, including: The groundwater level monitoring data is interpolated using ordinary kriging to obtain raster data with a preset spatial resolution, thus acquiring the first standard data. For multi-source feature variable data whose spatial resolution is not the preset spatial resolution, bilinear interpolation and majority class resampling are used to resample them to the preset spatial resolution to obtain second standard data. Based on the first and second standard data, a standard dataset with unified spatial resolution is obtained.

[0026] It is resampled to a preset spatial resolution using bilinear interpolation and majority class resampling, including: Bilinear interpolation is used to sample data from the meteorological data, remote sensing data, and nighttime light index that are not at the preset resolution to the preset spatial resolution. The majority class resampling method is used to sample data from the land use type that are not at the preset resolution to the preset spatial resolution.

[0027] Specifically, the meteorological data acquired includes precipitation, temperature, and potential evapotranspiration; the remote sensing data includes normalized vegetation index, surface temperature, surface elevation, and soil moisture content; and the human activity data includes land use type and nighttime light index. Ordinary Kriging interpolation was used to interpolate groundwater level data into 0.1°×0.1° raster data. This method is a geostatistical spatial estimation method that interpolates unsampled points by analyzing the spatial autocorrelation of known points. The majority class resampling method is used to standardize land use data. This method can process the spatial resolution of the data without destroying the initial classification and coding system, taking into account the characteristic that land use data must maintain integer coding. Other multi-source data are processed by bilinear interpolation for data spatial resolution standardization. This method uses the values ​​of the four adjacent grids around the point to be determined and calculates the new value by distance weighted averaging. This effectively eliminates the jagged edges generated when the image is magnified or resampled, making the data spatial distribution smoother and more continuous.

[0028] Furthermore, the normalization process for the standard dataset includes: ; in, For the normalized data, For initial data, and These are the maximum and minimum values ​​in the dataset, respectively.

[0029] Specifically, in S3, the groundwater level monitoring data and multi-source datasets are mapped to the [0,1] interval using the Min-Max normalization method, and divided into training set, validation set and test set according to the ratio of 0.7:0.1:0.2.

[0030] This method standardizes the data by mapping it to the [0,1] interval. The data is then divided into training, validation, and test sets in a ratio of 0.7:0.1:0.2. The training set is used for model parameter optimization, the validation set is used for model hyperparameter optimization, and the test set is used for the final performance evaluation of the model.

[0031] Furthermore, the Long Short-Term Memory network model includes: a bottom layer, an intermediate layer, a top layer, and a fully connected output layer; The underlying layer is used to extract the primary hidden state at the time step; The intermediate layer is used to randomly discard neurons in the primary hidden state with a preset probability; The top layer is used to extract high-level temporal features based on the discarded hidden state; The fully connected output layer is used to map the high-level temporal features into single-step water level prediction values.

[0032] Specifically, the constructed LSTM model includes a bottom LSTM, intermediate layers, a top LSTM, and a fully connected output layer. The bottom LSTM is responsible for outputting the initial feature sequence, the intermediate layers are responsible for randomly discarding data to prevent overfitting, and Dropout=0.2 is set, meaning that 20% of the data passed from the first layer to the second layer will be discarded. The top LSTM is responsible for further extracting features, and finally the output layer returns the prediction result.

[0033] Furthermore, the graph convolutional network model includes: a multi-factor adjacency matrix construction layer, a two-layer graph convolutional layer, and a fully connected output layer; The multi-factor adjacency matrix construction layer is used to obtain a comprehensive adjacency matrix by integrating three dimensions: geographical distance, elevation difference, and trend correlation. The dual-layer graph convolutional layer is used to perform first-order and second-order neighborhood information aggregation on the comprehensive adjacency matrix. The fully connected output layer is used to map the aggregated node features into single-step water level prediction values.

[0034] Specifically, the GCN model includes a multi-factor adjacency matrix construction, an input layer, a two-layer graph convolutional layer, and a fully connected output layer. The adjacency matrix integrates three dimensions: geographical distance, elevation difference, and trend correlation to obtain a comprehensive matrix. The input layer processes temporal features into spatial features. The first graph convolutional layer is responsible for collecting information from neighboring nodes for each node, and the second graph convolutional layer is responsible for collecting information from indirectly neighboring nodes for each node. The output layer returns the predicted value.

[0035] Furthermore, the Transformer model includes: an embedding layer, a positional encoding layer, a Transformer encoder, and an output layer; The embedding layer is used to linearly project multi-dimensional, multi-source features onto an embedding vector of uniform dimension. The location encoding layer is used to add learnable temporal location information to the embedding vector; The Transformer encoder is used to capture global temporal dependencies through a self-attention mechanism; The output layer is used to compress the encoder output into a single-step water level prediction value.

[0036] Specifically, the Transformer model includes an input embedding layer, a positional encoding layer, a Transformer encoder, and an output layer. The input embedding layer is responsible for inputting data and mapping the data to a unified feature space. The positional encoding layer provides the model with temporal information. The Transformer encoder can calculate the dot product similarity between time steps and generate an attention weight matrix based on the similarity. The output layer projects the abstract features back to the target space and outputs the final predicted value.

[0037] In detail, S4 constructs three deep learning models: Long Short-Term Memory Network (LSTM), Graph Convolutional Network (GCN), and Transformer model. All three models use the Adam optimizer, the loss function uses mean squared error, and an EarlyStopping callback mechanism is introduced to prevent overfitting.

[0038] Furthermore, obtaining the trained prediction model includes: The normalized data is input into the Long Short-Term Memory Network Model, the Graph Convolutional Network Model, and the Transformer Model. The Adam optimizer and the mean squared error loss function are used to iteratively update each model to obtain the trained Long Short-Term Memory Network Model, the Graph Convolutional Network Model, and the Transformer Model. Based on the trained Long Short-Term Memory (LSTM) network model, Graph Convolutional Network (GCN) model, and Transformer model, and combined with the multi-source feature variable dataset, we obtain the LSM network predicted water level, the Graph Convolutional Network predicted water level, and the Transformer model predicted water level, which correspond one-to-one with the spatial grid. Based on the water level predictions of the Long Short-Term Memory Network, Graph Convolutional Network, and Transformer Model, the mean absolute error of the Long Short-Term Memory Network, the mean absolute error of the Graph Convolutional Network, and the mean absolute error of the Transformer Model at each spatial grid are calculated. The model with the smallest mean absolute error at each spatial grid is marked as the optimal model for that grid. The optimal models of all grids are then spliced ​​together according to their spatial positions to form an optimal model combination covering the entire study area. This optimal model combination is then used as the trained prediction model.

[0039] Specifically, in S5, the mean absolute error (MAE) is used as an indicator to measure the accuracy of the prediction when evaluating the model's prediction performance, and then the optimal combination of models for groundwater level prediction is constructed.

[0040] Specifically, the mean absolute error (MAE) was chosen as the metric for evaluating the accuracy of groundwater level prediction for the three models. The formula for calculating this metric is as follows: ; in, For the true value, These are predicted values.

[0041] This index is used to evaluate the model prediction accuracy of each grid area, select the model with the smallest prediction error as the prediction model for that grid area, and integrate the prediction models of each grid area to select the optimal model combination for the entire area.

[0042] The following section uses the plains study area as a specific example: First, groundwater level monitoring data for the study area was obtained. In this case, the data platform of the Information Center of the Ministry of Water Resources was used to collect monthly groundwater level monitoring data from groundwater level monitoring stations in the study area from 2018 to 2024.

[0043] Ordinary Kriging interpolation was used to interpolate the groundwater level data into 0.1°×0.1° raster data. Taking the groundwater level in December 2024 as an example, the interpolation results are shown below. Figure 2 .

[0044] To build the model training set, it is necessary to further acquire meteorological data, remote sensing data, and human activity data.

[0045] Meteorological data includes: (1) Precipitation: Precipitation is the main source of replenishment for the groundwater system. Atmospheric precipitation infiltrates into the saturation zone through the vadose zone of the soil, directly driving the rise of the groundwater level.

[0046] (2) Temperature: Temperature mainly regulates groundwater dynamics indirectly through thermal effects. Rising temperatures significantly accelerate surface soil moisture evaporation and plant transpiration, leading to a drop in groundwater levels.

[0047] (3) Potential evapotranspiration: Potential evapotranspiration directly reflects the intensity of surface water evaporation and plant transpiration.

[0048] Remote sensing data includes: (1) Normalized Difference Vegetation Index (NDVI): NDVI is an important indicator reflecting vegetation cover density and growth vitality. The roots of vegetation absorb water, which in turn affects the groundwater level.

[0049] (2) Land surface temperature (LST): Surface temperature reflects soil heat flux, which directly drives the diffusion process of surface soil moisture.

[0050] (3) Surface elevation: Surface elevation determines the macroscopic hydraulic gradient of hydrogeological units, which has a significant impact on the flow path and recharge conditions of groundwater.

[0051] (4) Soil moisture content: Soil moisture content is a key link between surface processes and the groundwater system.

[0052] Human activity data includes: (1) Land use type: Different types of land use patterns directly affect the efficiency of surface water infiltration into the ground by altering the surface runoff coefficient and soil infiltration capacity.

[0053] (2) Nighttime light index: The nighttime light index represents the intensity of human activity and is generally positively correlated with the level of urbanization.

[0054] The majority class resampling method was used to process land use type data, reducing the spatial resolution of 1km×1km to 0.1°×0.1°. The processing results are shown below. Figure 3 .

[0055] For data with a spatial resolution other than 0.1°×0.1°, bilinear interpolation is used. Taking NDVI data and soil moisture content at a depth of 100-200cm as examples, bilinear interpolation is used to interpolate the 0.05°×0.05° NDVI data and the 0.25°×0.25° soil moisture content data into a dataset with a spatial resolution of 0.1°×0.1°. See details in [link to documentation]. Figure 4 and Figure 5 .

[0056] Subsequently, the groundwater level monitoring data and multi-source datasets were normalized using the Min-Max normalization method, mapping each data point to the [0,1] interval, and divided into training, validation, and test sets according to a ratio of 0.7:0.1:0.2.

[0057] We construct three deep learning models: LSTM, GCN, and Transformer.

[0058] For the LSTM model, firstly, its model layer structure is defined, including the bottom LSTM, intermediate layers, top LSTM, and fully connected output layer. In addition, building a complete model requires defining the global configuration and parameters, constructing data reading and feature extraction functions for data processing, constructing sequence construction functions to convert data into the three-dimensional tensors required by the LSTM, and finally defining the model training function and the result prediction and storage function.

[0059] The GCN model consists of a multi-factor adjacency matrix construction, an input layer, a two-layer graph convolutional layer, and a fully connected output layer. The adjacency matrix integrates three dimensions: geographical distance, elevation difference, and correlation of groundwater level change trends to obtain a comprehensive matrix. The input layer processes temporal features into spatial features. The first graph convolutional layer is responsible for collecting information from neighboring nodes for each node, and the second graph convolutional layer is responsible for collecting information from indirect neighboring nodes for each node. The output layer returns the predicted value. The Transformer model consists of an input embedding layer, a positional encoding layer, a Transformer encoder, and an output layer. The input embedding layer is responsible for taking in the data, expanding the data dimension, and mapping the data to a unified feature space. The positional encoding layer provides the model with temporal information. The Transformer encoder can calculate the dot product similarity between time steps and generate an attention weight matrix based on the similarity. The output layer compresses the high-dimensional features into numerical results and outputs the final predicted value.

[0060] The three models are trained and optimized by inputting training and validation set data. After training, the groundwater level of the test set is predicted and compared with the measured water level. The mean absolute error (MAE), coefficient of determination (R²), and root mean squared error (RMSE) were selected to evaluate the overall predictive performance of the model. The evaluation results are shown in Table 1.

[0061] Table 1 The LSTM model has the best overall prediction performance, with a lower mean absolute error than the other two models and a lower root mean square error than the Transformer model, but slightly higher than the GCN model. The GCN model has slightly better prediction performance than the Transformer model, but both its mean absolute error and root mean square error are lower than those of the Transformer model.

[0062] Based on the mean absolute error (MAE), the prediction model with the smallest error for each grid area is selected to form a combination of regional groundwater level prediction models. The MAEs of the LSTM, GCN, and Transformer models are shown in [reference needed]. Figure 6 , Figure 7 and Figure 8 The mean absolute error of the LSTM model is between 0.09 and 2.05; the mean absolute error of the GCN model is between 0.12 and 1.98; and the mean absolute error of the Transformer model is between 0.05 and 2.16.

[0063] The optimal model distribution is shown below. Figure 9 Based on this distribution, the training set, validation set, and test set are re-divided for each model; After completing model training and optimization, the optimal model combination was used to predict groundwater levels in each grid area of ​​the region. The evaluation indicators are shown in Table 2, and the distribution of the mean absolute error is shown in Table 3. Figure 10 The mean absolute error and root mean square error of the model group are both lower than the prediction results of the single model, further reducing the prediction error and effectively improving the prediction accuracy.

[0064] Table 2 The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A groundwater level prediction method based on a combination of multi-source data and a deep learning model, characterized in that, include: Acquire groundwater level monitoring data and multi-source characteristic variable data for the study area; Spatial resolution standardization was performed on the groundwater level monitoring data and multi-source characteristic variable data to obtain a standard dataset with uniform spatial resolution. The standard dataset is normalized, and a deep learning model is trained using the normalized data. The deep learning model includes a long short-term memory network model, a graph convolutional network model, and a Transformer model. The trained prediction model is used to predict the groundwater level at future times.

2. The groundwater level prediction method based on a combination of multi-source data and a deep learning model according to claim 1, characterized in that, The multi-source feature variable data includes: meteorological data, remote sensing data, and human activity data; The meteorological data includes: precipitation, temperature, and potential evapotranspiration; The remote sensing data includes: normalized vegetation index, surface temperature, surface elevation, and soil moisture content; The human activity data includes: land use type and nighttime light index.

3. The groundwater level prediction method based on a combination of multi-source data and a deep learning model according to claim 2, characterized in that, Spatial resolution standardization was performed on the groundwater level monitoring data and multi-source characteristic variable data to obtain a standard dataset with unified spatial resolution, including: The groundwater level monitoring data is interpolated using ordinary kriging to obtain raster data with a preset spatial resolution, thus acquiring the first standard data. For multi-source feature variable data whose spatial resolution is not the preset spatial resolution, bilinear interpolation and majority class resampling are used to resample them to the preset spatial resolution to obtain second standard data. Based on the first and second standard data, a standard dataset with unified spatial resolution is obtained.

4. The groundwater level prediction method based on a combination of multi-source data and a deep learning model according to claim 3, characterized in that, It is resampled to a preset spatial resolution using bilinear interpolation and majority class resampling, including: Bilinear interpolation is used to sample data from the meteorological data, remote sensing data, and nighttime light index that are not at the preset resolution to the preset spatial resolution. The majority class resampling method is used to sample data from the land use type that are not at the preset resolution to the preset spatial resolution.

5. The groundwater level prediction method based on a combination of multi-source data and a deep learning model according to claim 1, characterized in that, Normalizing the standard dataset includes: ； in, For the normalized data, For initial data, and These are the maximum and minimum values ​​in the dataset, respectively.

6. The groundwater level prediction method based on a combination of multi-source data and a deep learning model according to claim 1, characterized in that, The Long Short-Term Memory (LSTM) network model includes: a bottom layer, a middle layer, a top layer, and a fully connected output layer; The underlying layer is used to extract the primary hidden state at the time step; The intermediate layer is used to randomly discard neurons in the primary hidden state with a preset probability; The top layer is used to extract high-level temporal features based on the discarded hidden state; The fully connected output layer is used to map the high-level temporal features into single-step water level prediction values.

7. The groundwater level prediction method based on a combination of multi-source data and a deep learning model according to claim 1, characterized in that, The graph convolutional network model includes: a multi-factor adjacency matrix construction layer, a two-layer graph convolutional layer, and a fully connected output layer; The multi-factor adjacency matrix construction layer is used to obtain a comprehensive adjacency matrix by integrating three dimensions: geographical distance, elevation difference, and trend correlation. The dual-layer graph convolutional layer is used to perform first-order and second-order neighborhood information aggregation on the comprehensive adjacency matrix. The fully connected output layer is used to map the aggregated node features into single-step water level prediction values.

8. The groundwater level prediction method based on a combination of multi-source data and a deep learning model according to claim 1, characterized in that, The Transformer model includes: an embedding layer, a positional encoding layer, a Transformer encoder, and an output layer; The embedding layer is used to linearly project multi-dimensional, multi-source features onto an embedding vector of uniform dimension. The location encoding layer is used to add learnable temporal location information to the embedding vector; The Transformer encoder is used to capture global temporal dependencies through a self-attention mechanism; The output layer is used to compress the encoder output into a single-step water level prediction value.

9. The groundwater level prediction method based on a combination of multi-source data and a deep learning model according to claim 1, characterized in that, Obtaining a trained prediction model includes: The normalized data is input into the Long Short-Term Memory Network Model, the Graph Convolutional Network Model, and the Transformer Model. The Adam optimizer and the mean squared error loss function are used to iteratively update each model to obtain the trained Long Short-Term Memory Network Model, the Graph Convolutional Network Model, and the Transformer Model. Based on the trained Long Short-Term Memory (LSTM) network model, Graph Convolutional Network (GCN) model, and Transformer model, and combined with the multi-source feature variable dataset, we obtain the LSM network predicted water level, the Graph Convolutional Network predicted water level, and the Transformer model predicted water level, which correspond one-to-one with the spatial grid. Based on the water level predictions of the Long Short-Term Memory Network, Graph Convolutional Network, and Transformer Model, the mean absolute error of the Long Short-Term Memory Network, the mean absolute error of the Graph Convolutional Network, and the mean absolute error of the Transformer Model at each spatial grid are calculated. The model with the smallest mean absolute error at each spatial grid is marked as the optimal model for that grid. The optimal models of all grids are then spliced ​​together according to their spatial positions to form an optimal model combination covering the entire study area. This optimal model combination is then used as the trained prediction model.

Images