Training method of underground water level prediction model and related device

Abstract

The invention discloses an underground water level prediction model training method and a related device, and the method comprises the steps: determining underground water level sample data of a plurality of monitoring points, carrying out the feature extraction of the underground water level sample data of the plurality of monitoring points through a spatial feature extraction module, and obtaining a spatial feature vector; and processing the spatial feature vector through a time feature extraction module to obtain a spatio-temporal feature vector. And according to the spatial-temporal feature vectors, performing prediction through the initial prediction model to obtain predicted underground water levels of the monitoring points. According to the difference between the predicted groundwater level of each monitoring point and the groundwater level label, adjusting model parameters of the initial prediction model to obtain a groundwater level prediction model so that the groundwater level prediction model can be used for predicting the groundwater levels of the multiple monitoring points. Therefore, the model recognizes the influence of water level fluctuation on the underground water level corresponding to a certain monitoring point at different positions with spatial correlation, and the accuracy of predicting the underground water level by the model is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and particularly relates to a method for training an underground water level prediction model and related devices. Background Art

[0002] With the implementation of ecological water replenishment work, the underground water levels in some plain areas have been continuously rising. However, while the rising of the underground water level brings ecological restoration benefits, it may also affect the safety of underground projects, resulting in cracking, leakage, seepage or even gushing water of underground structures, and weakening the bearing capacity of the foundation, thereby increasing the difficulty of waterproofing and anti-floating of underground projects. Therefore, it is particularly important to predict the underground water level in a timely and accurate manner.

[0003] In related technologies, a numerical simulation model is usually used to predict the underground water level. This method is established based on the physical process of groundwater flow. On the basis of generalizing the aquifer structure, boundaries, source-sink terms, etc., and with the help of mathematical equations, the underground water level is calculated. In order to make up for the insufficient ability of the numerical simulation model to handle non-linear relationships, related technologies also use deep learning models such as recurrent neural networks to predict the underground water level.

[0004] However, the above methods only rely on historical laws and do not consider the spatial influence of the underground water level by the surrounding water level fluctuations, resulting in low accuracy of the predicted underground water level. Summary of the Invention

[0005] In view of the above problems, the present application provides a method for training an underground water level prediction model and related devices to improve the accuracy of predicting the underground water level.

[0006] Based on this, the present application discloses the following technical solutions:

[0007] In a first aspect, an embodiment of the present application provides a method for training an underground water level prediction model, the method including:

[0008] Determine the underground water level sample data of multiple monitoring points, where the multiple monitoring points are multiple monitoring points from the same area and having spatial correlation, and the underground water level sample data includes an underground water level label, hydrogeological feature data of the monitoring points, and spatial feature data, and the spatial feature data is used to indicate the spatial correlation relationship between the multiple monitoring points;

[0009] Extract features from the underground water level sample data of the multiple monitoring points through a spatial feature extraction module to obtain a spatial feature vector, where the spatial feature extraction module is used to capture the spatial correlation relationship between different monitoring points;

[0010] The time feature extraction module processes the spatial feature vector to obtain a spatio-temporal feature vector, and the time feature extraction module is used to capture the variation relationship of the groundwater level over time series;

[0011] According to the spatio-temporal feature vector, prediction is performed through an initial prediction model to obtain the predicted groundwater levels of each of the monitoring points;

[0012] According to the differences between the predicted groundwater levels of each of the monitoring points and the groundwater level labels, the model parameters of the initial prediction model are adjusted to obtain the groundwater level prediction model, so that the groundwater level prediction model is used to predict the groundwater levels of multiple monitoring points.

[0013] Optionally, the method further includes:

[0014] Obtain the historical groundwater level data of multiple monitoring points, where the historical groundwater level data is used to indicate the groundwater level state information from the (i - n)-th period to the (i - 1)-th period in the past, n is a positive integer greater than 1 and less than i, i is a positive integer, and the (i - n)-th period to the (i - 1)-th period are multiple consecutive periods distributed in chronological order;

[0015] Predict the historical groundwater level data through the groundwater level prediction model to obtain the first groundwater level in the future i-th period;

[0016] Predict the historical groundwater level data and the first groundwater level through the groundwater level prediction model to obtain the second groundwater level in the future (i + 1)-th period.

[0017] Optionally, the spatial feature data is determined in the following manner:

[0018] Obtain the geographical distribution information of the multiple monitoring points;

[0019] According to the geographical distribution information, determine the spatial association relationship between any two monitoring points in the area to obtain a plurality of spatial association relationships;

[0020] According to the plurality of spatial association relationships, construct an adjacency matrix of the multiple monitoring points to obtain the spatial feature data.

[0021] Optionally, the multiple monitoring points include a first monitoring point and a second monitoring point, and the determining the spatial association relationship between any two monitoring points in the area according to the geographical distribution information to obtain a plurality of spatial association relationships includes:

[0022] According to the geographical distribution information, determine the spatial distance between the first monitoring point and the second monitoring point;

[0023] Determine the spatial association weight between the first monitoring point and the second monitoring point according to the spatial distance, where the spatial relationship weight is used to identify the spatial association strength between the monitoring points, and the smaller the spatial distance, the greater the spatial association weight;

[0024] Determine the spatial association relationship between the first monitoring point and the second monitoring point according to the spatial association weight;

[0025] Take any two monitoring points as the first monitoring point and the second monitoring point respectively to obtain the multiple spatial association relationships.

[0026] Optionally, the spatial feature extraction module includes a graph convolutional neural network and a spatial attention mechanism. Feature extraction of the groundwater level sample data of the multiple monitoring points through the spatial feature extraction module to obtain a spatial feature vector, including:

[0027] According to the hydrogeological feature data and the spatial feature data of the monitoring point, perform feature extraction through the graph convolutional neural network to obtain an initial spatial feature vector;

[0028] Generate different first weights for the features of different categories in the initial spatial feature vector through the spatial attention mechanism;

[0029] Determine the spatial feature vector according to the initial spatial feature vector and the first weights corresponding to the features of different categories in the initial spatial feature vector.

[0030] Optionally, the time feature extraction module includes a recurrent neural network and a time attention mechanism. Feature extraction of the spatial feature vector through the time feature extraction module to obtain a spatio-temporal feature vector, including:

[0031] Process the spatial feature vector through the recurrent neural network to obtain the initial spatio-temporal feature vector;

[0032] Generate different second weights for the features of different time periods in the initial spatio-temporal feature vector through the time attention mechanism;

[0033] Determine the spatio-temporal feature vector according to the initial spatio-temporal feature vector and the second weights corresponding to different time periods in the initial spatio-temporal feature vector.

[0034] Optionally, the hydrogeological feature data includes dynamic sample data and static sample data. The dynamic sample data is data that changes with time periods, including groundwater level, precipitation, extraction volume, and ecological recharge volume. The static sample data is data that does not change with time periods, including longitude and latitude, digital elevation value, rainfall infiltration coefficient, and land use rate.

[0035] In a second aspect, an embodiment of the present application provides a training device for an underground water level prediction model. The device includes a determination unit, a first processing unit, a second processing unit, a prediction unit, and an adjustment unit;

[0036] The determination unit is configured to determine the underground water level sample data of a plurality of monitoring points. The plurality of monitoring points are a plurality of monitoring points from the same area and having spatial correlation. The underground water level sample data includes an underground water level label, the hydrogeological characteristic data of the monitoring points, and spatial characteristic data. The spatial characteristic data is used to indicate the spatial correlation relationship between the plurality of monitoring points;

[0037] The first processing unit is configured to perform feature extraction on the underground water level sample data of the plurality of monitoring points through a spatial feature extraction module to obtain a spatial feature vector. The spatial feature extraction module is used to capture the spatial correlation relationship between different monitoring points;

[0038] The second processing unit is configured to process the spatial feature vector through a time feature extraction module to obtain a spatio-temporal feature vector. The time feature extraction module is used to capture the change relationship of the underground water level over time series;

[0039] The prediction unit is configured to perform prediction through an initial prediction model according to the spatio-temporal feature vector to obtain the predicted underground water level of each monitoring point;

[0040] The adjustment unit is configured to adjust the model parameters of the initial prediction model according to the difference between the predicted underground water level and the underground water level label of each monitoring point to obtain the underground water level prediction model, so that the underground water level prediction model is used to predict the underground water level of a plurality of monitoring points.

[0041] In a third aspect, an embodiment of the present application provides a computer device. The computer device includes a processor and a memory:

[0042] The memory is used to store a computer program and transmit the computer program to the processor;

[0043] The processor is configured to execute the method described in the first aspect above according to the computer program.

[0044] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium. The computer-readable storage medium is used to store a computer program. The computer program is used to execute the method described in the first aspect above.

[0045] Fifth aspect, an embodiment of the present application provides a computer program product including a computer program, which, when running on a computer device, causes the computer device to execute the method described in the above first aspect.

[0046] As can be seen from the above technical solutions, the present application has at least the following beneficial effects:

[0047] The present application no longer solely relies on the historical laws of sample data, but also captures the spatial correlation relationships between multiple monitoring points on the basis of historical laws, so as to train a groundwater level prediction model with higher prediction accuracy.

[0048] For multiple monitoring points from the same region and having spatial correlations, determine the groundwater level sample data of the multiple monitoring points, where the groundwater level sample data includes groundwater level labels, hydrogeological characteristic data of the monitoring points, and spatial characteristic data. Since the hydrogeological characteristics of different monitoring points are correlated with each other in the spatial dimension, the groundwater level corresponding to each monitoring point is affected by the water level fluctuations at other locations. Therefore, by determining the spatial characteristic data used to indicate the spatial correlation relationships between multiple monitoring points, a data source in the spatial dimension can be provided for training the model.

[0049] Extract features from the groundwater level sample data of multiple monitoring points through a spatial feature extraction module to obtain spatial feature vectors, so as to capture the spatial correlation relationships between different monitoring points. Process the spatial feature vectors through a time feature extraction module to obtain spatio-temporal feature vectors, so as to further capture the variation relationship of the groundwater level over time series on the basis of the spatial feature vectors. According to the spatio-temporal feature vectors, perform predictions through an initial prediction model to obtain the predicted groundwater levels of each monitoring point. Adjust the model parameters of the initial prediction model according to the differences between the predicted groundwater levels of each monitoring point and the groundwater level labels to obtain a groundwater level prediction model for predicting the groundwater levels of multiple monitoring points. Thus, by first determining the groundwater level sample data that can reflect the spatial correlation relationships between different monitoring points and capturing this correlation relationship through a spatial feature extraction module, during the process of training the model, it is not only possible to predict the future groundwater level based on the historical laws presented by the groundwater level in the past period, but also possible to make the model recognize the influence of the water level fluctuations at different locations with spatial correlations on the groundwater level corresponding to a certain monitoring point by adjusting the model parameters, so as to obtain a groundwater level prediction model with the ability of spatial correlation analysis, improving the accuracy of the model in predicting the groundwater level. Description of the Drawings

[0050] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments recorded in the present application. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0051] Figure 1 Flow diagram of a method for training an underground water level prediction model provided by an embodiment of the present application;

[0052] Figure 2 Flow diagram of a spatial attention mechanism provided by an embodiment of the present application;

[0053] Figure 3 Flow diagram of a temporal attention mechanism provided by an embodiment of the present application;

[0054] Figure 4 Flow diagram of a method for training an underground water level prediction model provided by an embodiment of the present application;

[0055] Figure 5 Flow diagram of a training technical architecture provided by an embodiment of the present application;

[0056] Figure 6 Schematic diagram of a training principle provided by an embodiment of the present application;

[0057] Figure 7 Schematic diagram of the structure of a training device for an underground water level prediction model provided by an embodiment of the present application;

[0058] Figure 8 Schematic diagram of the structure of a computer device provided by an embodiment of the present application. Detailed implementation manners

[0059] The following will describe the embodiments of the present application in more detail with reference to the accompanying drawings. Although some embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. On the contrary, these embodiments are provided to more thoroughly and completely understand the present application. It should be understood that the accompanying drawings and embodiments of the present application are only for exemplary purposes and are not used to limit the protection scope of the present application.

[0060] In the related art, due to the fact that the groundwater level is often non-linear affected by precipitation, ecological water replenishment, exploitation, boundary conditions, etc., and combined with the characteristics of the heterogeneity of the aquifer and the lag of recharge, it is difficult to generalize the numerical simulation model and identify parameters. The deep learning model has certain advantages in solving the above non-linear problems. It ignores the physical mechanism of groundwater flow and focuses on exploring the relationship between the influencing factors of the groundwater level and its dynamics.

[0061] Taking the long short-term memory network (LSTM, Long Short-Term Memory) as an example, because it can effectively process time series data and has the function of long-term memory, LSTM can be used for the prediction of surface and groundwater levels. Due to the problems of information loss and gradient disappearance in the processing of long sequence data by traditional LSTM, it is difficult to ensure the robustness of the prediction. In recent years, the attention mechanism has been gradually introduced into the LSTM model to solve such problems. The dynamic changes of the groundwater level are complex, not only relying on historical laws, but also affected by the spatial influence of the water level fluctuations around. Simply relying on LSTM is difficult to fully capture the spatial correlation, resulting in a low accuracy of predicting the groundwater level by LSTM.

[0062] Based on this, the embodiments of the present application provide a training method and related device for a groundwater level prediction model. By first determining the groundwater level sample data that can reflect the spatial correlation relationship between different monitoring points, and capturing this correlation relationship through a spatial feature extraction module, so that in the process of training the model, not only can the future groundwater level be predicted based on the historical laws presented by the groundwater level in the past period, but also by adjusting the model parameters, the model can identify the influence of the water level fluctuations at different positions with spatial correlation on the corresponding groundwater level of a certain monitoring point, thereby obtaining a groundwater level prediction model with the ability of spatial correlation analysis and improving the accuracy of the model in predicting the groundwater level.

[0063] The training method of the groundwater level prediction model provided by the present application can be applied to computer devices with model training capabilities, such as terminal devices and servers. Among them, the terminal device can specifically be a desktop computer, a laptop computer, a mobile phone, a tablet computer, etc.; the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, etc. The terminal device and the server can be directly or indirectly connected through wired or wireless communication methods, and the present application does not limit this.

[0064] See Figure 1 , this figure is a schematic flowchart of the training method of the groundwater level prediction model provided by the embodiments of the present application. For the convenience of description, the following embodiments are introduced by taking the execution subject of the training method of the groundwater level prediction model as a server as an example. As Figure 1As shown in the figure, the training method of the groundwater level prediction model includes S101 - S105.

[0065] S101: Determine the groundwater level sample data of multiple monitoring points.

[0066] The monitoring point is a specific location selected within a specific area for collecting groundwater level and related hydrogeological characteristic data. For example, the monitoring point can be a groundwater observation well, and the groundwater level sample data is the sample data used to train the groundwater level prediction model. As an implementation method, the groundwater level can use the groundwater depth as an evaluation index.

[0067] Among them, the multiple monitoring points are multiple monitoring points from the same area and having spatial correlation. The groundwater system usually does not exist in isolation but is a complex interconnected system. The groundwater level change in one area often affects its neighboring areas. For example, if a large amount of rainfall or ecological water replenishment occurs in the area near a certain monitoring point, this may cause the groundwater level in this area to rise, and this water level rise may affect the groundwater level of adjacent monitoring points. Therefore, for a certain monitoring point, if its hydrogeological conditions change, it will cause fluctuations in the groundwater level, thereby affecting the groundwater levels corresponding to multiple monitoring points having spatial correlation with this monitoring point.

[0068] The groundwater level sample data includes groundwater level labels, hydrogeological characteristic data of the monitoring points, and spatial characteristic data. The groundwater level labels are used to identify the true groundwater levels collected at the monitoring points. The hydrogeological characteristic data is the data reflecting the hydrogeological conditions in the area where the monitoring points are located. During the model training process, the hydrogeological characteristic data is used to help the model understand the influence of the hydrogeological characteristics of each monitoring point on the groundwater level. The spatial characteristic data is used to indicate the spatial correlation relationship between multiple monitoring points.

[0069] As can be seen from the foregoing, in the related art, the spatial correlation of different groundwater levels is not considered. Therefore, during the process of training the model, the sample data also does not have data for characterizing the spatial correlation relationship at different positions. In this application, not only can the hydrogeological characteristic data and the groundwater level labels for identifying the true groundwater levels be obtained through the detection devices at the monitoring points, but also the spatial characteristic data can be determined according to the spatial correlation relationship between each monitoring point, so that during the process of training the groundwater level prediction model, the model can learn the spatial correlation relationship between different monitoring points, improving the comprehensiveness of the sample data.

[0070] In a possible implementation method, the hydrogeological characteristic data includes dynamic sample data and static sample data.

[0071] Among them, the dynamic sample data are data that change with time periods, including groundwater levels, precipitation, extraction amounts, and ecological water replenishment amounts. For example, in a daily-scale time period, the dynamic sample data include the groundwater levels, extraction amounts, and ecological water replenishment amounts corresponding to each day respectively.

[0072] The static sample data are data that do not change with time periods, including longitude and latitude, digital elevation values. The digital elevation value is the value corresponding to the Digital Elevation Model (DEM) at the monitoring point, rainfall infiltration coefficient, and land use rate. The static sample data are the inherent hydrogeological properties of the monitoring point.

[0073] Thus, the model can not only analyze the inherent hydrogeological characteristic data of the monitoring point itself through the static sample data, but also enable the model to capture the change trends of the groundwater level-related data at different time periods during the training process through the dynamic sample data, which is more conducive to capturing the historical laws of the groundwater level through the time feature extraction module, and improves the accuracy of predicting the groundwater level.

[0074] As an implementation method, the time sliding window method can be adopted to extract the hydrogeological characteristic data of multiple time periods (such as groundwater level, rainfall, extraction amount, and ecological water replenishment, etc.) as the groundwater level sample data. For example, taking the daily scale as the time period and 7 days as a window, the hydrogeological characteristic data for 7 consecutive days are obtained, and then the groundwater level on the 8th day is used as the groundwater level label. The groundwater level corresponding to the last day of the time window is used as the output target. This time window division method can capture the short-term dynamic changes of the input variables and improve the model's learning ability for time series features.

[0075] S102: Extract features from the groundwater level sample data of multiple monitoring points through the spatial feature extraction module to obtain spatial feature vectors.

[0076] The initial prediction model is a groundwater level prediction model that has not been trained yet. The initial prediction model has the same model structure as the groundwater level prediction model, but different model parameters. By adjusting the model parameters, the prediction accuracy of the initial prediction model can be gradually improved. This spatial feature extraction module is a component of the model structure of the initial prediction model, which is used to capture the spatial correlation relationships between different monitoring points. Taking the groundwater level sample data as the input, the spatial correlation factors such as geographical location and geological conditions between each monitoring point can be identified through the spatial feature extraction module, and then encoded to obtain spatial feature vectors. The spatial feature vector is a set of numerical values obtained by processing the groundwater level sample data through the spatial feature extraction module, which can reflect the spatial correlation relationships between the monitoring points.

[0077] As an implementation, since there are outliers, duplicate values, and missing values in the original sample data, which directly affect the model training and prediction accuracy, before the groundwater level sample data is input into the spatial feature extraction module, the accuracy of the sample data can be improved through preprocessing. For example, first, by setting reasonable thresholds and examining the data distribution, outliers and duplicate records are removed. Then, the cubic spline interpolation method is used to fill in the missing data to ensure the integrity of the time series data. Finally, wavelet transform is used to eliminate data noise and improve the accuracy and reliability of the data.

[0078] S103: Process the spatial feature vector through the time feature extraction module to obtain the spatio-temporal feature vector.

[0079] The time feature extraction module is a component of the initial prediction model's structure, which is used to capture the variation relationship of the groundwater level over time. Thus, the spatial feature vector that can reflect the spatial correlation relationship of the monitoring points is input into the time feature extraction module. The time feature extraction module identifies the influence of the past-period groundwater level sample data on the future groundwater level, and then encodes the spatial feature vector to obtain the spatio-temporal feature vector. The spatio-temporal feature vector is a set of numerical values obtained by processing the spatial feature vector through the time feature extraction module, which not only has the feature representation of the spatial correlation relationship of the monitoring points but also can reflect the dynamic change of the groundwater level over time periods.

[0080] S104: According to the spatio-temporal feature vector, perform prediction through the initial prediction model to obtain the predicted groundwater levels of each monitoring point.

[0081] The predicted groundwater level is the result of the initial prediction model predicting the groundwater level based on the spatio-temporal feature data.

[0082] S105: According to the difference between the predicted groundwater levels of each monitoring point and the groundwater level labels, adjust the model parameters of the initial prediction model to obtain the groundwater level prediction model, so that the groundwater level prediction model can be used to predict the groundwater levels of multiple monitoring points.

[0083] According to the difference between the predicted groundwater level and the groundwater level label, adjust the model parameters of the initial prediction model so that the difference between the predicted groundwater level and the groundwater level label becomes smaller and smaller, such as meeting the number of iterations or model convergence, etc. This application does not make specific limitations on this. Thus, by reducing the difference between the predicted groundwater level and the groundwater level label, the prediction accuracy of the initial prediction model is continuously improved, and thus a groundwater level prediction model with higher prediction accuracy is obtained.

[0084] For example, during the training process, the model performance is optimized by adjusting hyperparameters (such as the number of iterations, batch size, number of neurons, etc.) until the loss function converges to the minimum. To prevent the model from overfitting, the following strategies are introduced during the training process: adding a Dropout layer to randomly discard some neurons in the model to reduce the dependence on specific weights; constraining the weight size through L2 regularization to prevent the model from overfitting the training data. To further improve the optimization effect, a learning rate scheduler (with an initial learning rate of 0.001, and if the error does not decrease after 5 training iterations, the learning rate automatically becomes half of the original) is used to dynamically adjust the learning rate of the Adaptive Moment Estimation (Adam) optimizer, enabling the model to decline rapidly in the initial stage of convergence and tend to be stable in the later stage. The following three metrics are used to evaluate the prediction accuracy of the model: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Nash-Sutcliffe Efficiency (NSE). If the evaluation results meet the accuracy requirements, the model training is completed; otherwise, continue to optimize. See Formulas 1-3:

[0085]

[0086] Among them, n represents the total number of predictions, Y represents the predicted groundwater level, X represents the groundwater level samples, represents the average value of X, and i is used to identify the i-th time period.

[0087] It can be seen from the above technical solutions that this application no longer solely relies on the historical patterns of sample data, but also captures the spatial correlation relationships between multiple monitoring points based on the historical patterns to train a groundwater level prediction model with higher prediction accuracy.

[0088] For multiple monitoring points from the same area and with spatial correlation, determine the groundwater level sample data of the multiple monitoring points, where the groundwater level sample data includes groundwater level labels, hydrogeological characteristic data of the monitoring points, and spatial characteristic data. Since the hydrogeological characteristics of different monitoring points are spatially correlated with each other, the groundwater level corresponding to each monitoring point is affected by the water level fluctuations at other locations. Therefore, by determining the spatial characteristic data used to indicate the spatial correlation relationships between multiple monitoring points, a data source in the spatial dimension can be provided for training the model.

[0089] The spatial feature extraction module extracts features from the groundwater level sample data of multiple monitoring points to obtain spatial feature vectors, thereby being able to capture the spatial correlation relationships between different monitoring points. The temporal feature extraction module processes the spatial feature vectors to obtain spatio-temporal feature vectors, thereby being able to further capture the variation relationships of the groundwater level over time series based on the spatial feature vectors. According to the spatio-temporal feature vectors, predictions are made through the initial prediction model to obtain the predicted groundwater levels of each monitoring point. According to the differences between the predicted groundwater levels of each monitoring point and the groundwater level labels, the model parameters of the initial prediction model are adjusted to obtain the groundwater level prediction model, so that the groundwater level prediction model can be used to predict the groundwater levels of multiple monitoring points. Thus, by first determining the groundwater level sample data that can reflect the spatial correlation relationships between different monitoring points and capturing this correlation relationship through the spatial feature extraction module, not only can the future groundwater levels be predicted based on the historical patterns presented by the groundwater levels in the past periods during the process of training the model, but also the model can be adjusted to recognize the impacts of the water level fluctuations at different locations with spatial correlations on the corresponding groundwater levels of a certain monitoring point, thereby obtaining a groundwater level prediction model with the ability of spatial correlation analysis and improving the accuracy of the model in predicting the groundwater levels.

[0090] In a possible implementation manner, the trained groundwater level prediction model can be used to predict the groundwater levels of multiple future periods in a rolling prediction manner.

[0091] The historical groundwater level data of multiple monitoring points is obtained. The historical groundwater level data is used to indicate the groundwater level state information from the (i - n)-th period to the (i - 1)-th period in the past, where n is a positive integer greater than 1 and less than i, and i is a positive integer. The (i - n)-th period to the (i - 1)-th period are consecutive multiple periods distributed in chronological order. For example, the historical groundwater level data is the rainfall, extraction volume, and ecological replenishment volume corresponding to each monitoring point in the past 7 days, so as to simulate the hydrogeological state changes of the groundwater system through the above-mentioned relevant state information of the groundwater level and use it as input data to be input into the groundwater level prediction model. Another example is that the historical groundwater level data can also include the real groundwater levels of each monitoring point in the past 7 days, so as to enhance the prediction accuracy of the future groundwater levels based on the historical patterns of the real groundwater levels.

[0092] The groundwater level prediction model is used to predict the historical groundwater level data to obtain the first groundwater level of the future i-th period.

[0093] The i-th time period is a future time period after the (i - 1)-th time period. From the (i - n)-th time period to the i-th time period are multiple consecutive time periods in terms of time. The first groundwater level is the result of predicting the groundwater level of the i-th time period by the groundwater level prediction model based on the historical groundwater level data. Thus, the groundwater level of the nearest future time period can be predicted by the groundwater level prediction model based on the historical groundwater level data of multiple past time periods.

[0094] The groundwater level prediction model predicts the historical groundwater level data and the first groundwater level to obtain the second groundwater level of the (i + 1)-th future time period.

[0095] The second groundwater level is the result of predicting the groundwater level of the (i + 1)-th time period by the groundwater level prediction model based on the historical groundwater level data. The (i + 1)-th time period is a future time period after the i-th time period.

[0096] After predicting the first groundwater level of the i-th time period, the historical groundwater level data and the first groundwater level can be combined as the input data for predicting the groundwater level of the (i + 1)-th time period, that is, the prediction result of a certain time period is used as the input for predicting the next time period, and the prediction results of multiple future time periods are gradually generated by recursion.

[0097] Thus, in the case of predicting the groundwater level of one time period, by adopting the recursive prediction method, the groundwater levels of multiple future time periods can be predicted. The prediction results of multiple time periods are not only based on historical laws but also consider the spatial correlation relationship between monitoring points, so that the groundwater levels of more time periods can be accurately predicted.

[0098] In a possible implementation manner, the spatial feature data is determined as follows:

[0099] Obtain the geographical distribution information of multiple monitoring points. The geographical distribution information is the position distribution information of the monitoring points in the geographical dimension, such as the longitude and latitude coordinates of the monitoring points, digital elevation values, etc.

[0100] According to the geographical distribution information, determine the spatial correlation relationship between any two monitoring points in the area to obtain multiple spatial correlation relationships. Through the position relationship indicated by the geographical distribution information, the correlation relationship between any two monitoring points in the spatial dimension can be determined. This correlation relationship is determined based on the geographical location distribution of the monitoring points. For example, the target monitoring point has a spatial correlation relationship with its surrounding monitoring points. Among them, a pair of monitoring points determines a corresponding spatial correlation relationship, so that multiple monitoring points can determine multiple spatial correlation relationships.

[0101] Specifically, taking the first monitoring point and the second monitoring point included in multiple monitoring points as an example, an exemplary introduction to a method for determining the spatial correlation relationship is given. See A1 - A.

[0102] A1: First, determine the spatial distance between the first monitoring point and the second monitoring point according to the geographical distribution information. For example, the spatial distance is calculated through the geographical coordinates of the two monitoring points.

[0103] A2: Determine the spatial correlation weight between the first monitoring point and the second monitoring point according to the spatial distance. The spatial relationship weight is used to identify the strength of the spatial correlation between the monitoring points.

[0104] Among them, the smaller the spatial distance, the larger the spatial correlation weight. That is to say, if the second monitoring point is closer to the first monitoring point, the influence of the water level fluctuation of the second monitoring point on the groundwater level of the first monitoring point is greater. For example, the spatial correlation weight can be determined through the following formula:

[0105]

[0106] where ee i,,j is the spatial correlation weight, ll i,，j is the distance between the first monitoring point and the second monitoring point, and LL is the average distance between each monitoring point.

[0107] A3: Determine the spatial correlation relationship between the first monitoring point and the second monitoring point according to the spatial correlation weight. For example, the spatial correlation relationship between the two monitoring points can be established through an edge, and the spatial correlation relationship is used as the weight of the edge.

[0108] A4: Finally, take any two monitoring points as the first monitoring point and the second monitoring point respectively, and execute the above A1 - A3 to obtain multiple spatial correlation relationships.

[0109] Thus, the spatial correlation strength of monitoring points at different distances can be quantified through the spatial correlation weight, so that the model can pay more attention to the sample data of the monitoring points near the monitoring point. The sample data of these adjacent monitoring points has a stronger influence on the groundwater level in space, thereby improving the prediction accuracy of the model.

[0110] Construct an adjacency matrix of multiple monitoring points based on multiple spatial correlation relationships to obtain spatial feature data. The adjacency matrix is a data structure representing the spatial correlation relationship between each monitoring point. Each element in the adjacency matrix is used to characterize the spatial correlation relationship between two monitoring points. For example, whether two monitoring points have a spatial correlation relationship is represented by 1 and 0. Another example is that the specific weight of 0 - 1 can be used to determine the spatial correlation strength between two monitoring points.

[0111] The adjacency matrix abstracts the spatial correlation relationship between each monitoring point in the form of an element lattice, which is easy for the model to understand. Thus, it can more accurately characterize the spatial correlation between each monitoring point, enabling the model to more accurately learn the spatial correlation relationship reflected by the spatial feature data.

[0112] In a possible implementation, the spatial feature extraction module includes a graph convolutional neural network (GCN) and a spatial attention mechanism. The graph convolutional neural network is a neural network for processing graph-structured data (such as spatial feature data). The spatial attention mechanism is a technique that allows the model to assign weights to different categories of hydrogeological feature data.

[0113] In the process of obtaining the spatial feature vector, feature extraction can be performed through the graph convolutional neural network based on the hydrogeological feature data and spatial feature data of the monitoring points to obtain an initial spatial feature vector. The initial spatial feature vector is a preliminary feature representation calculated by the graph convolutional neural network based on the hydrogeological feature data and spatial feature data of the monitoring points.

[0114] Through the spatial attention mechanism, different first weights are generated for the features of different categories in the initial spatial feature vector. As described above, the hydrogeological feature data includes multiple categories, such as precipitation, extraction volume, and ecological recharge volume, etc. However, the influence degrees of different categories on the groundwater level are different. The spatial attention mechanism can assign corresponding first weights to different categories of hydrogeological feature data, so as to measure the influence degree of each category on the groundwater level through the first weights.

[0115] Then, according to the initial spatial feature vector and the first weights corresponding to the features of different categories in the initial spatial feature vector, the spatial feature vector is determined.

[0116] It should be noted that, in the case of unified feature vector dimensions, the spatial attention mechanism provided by the embodiments of the present application can also be used in combination with the temporal attention mechanism. For example, for the feature vector output by the recurrent neural network, the spatial feature attention mechanism assigns first weights to different categories of hydrogeological feature data and generates a spatial feature vector, and then the temporal attention mechanism processes the spatial feature vector. The embodiments of the present application do not make any limitations here.

[0117] For example, referring to Figure 2 Formulas 5 and 6, the initial spatial feature vector is divided for each time period to form t vectors m is the number of input features, t is the number of input time steps (time periods). The attention value of the importance of each feature to the groundwater level is calculated through the Sigmoid activation function, and the attention weights α_t of each feature are generated through Softmax normalization (Formula 5) to ensure that the sum of the attention weights of all input variables at each sampling moment t is 1. Finally, the weighted sample of the spatial attention is obtained through the element-wise multiplication of the two vectors (Formula 6).

[0118]

[0119] Thus, after extracting features from the spatial correlation relationship represented by the groundwater level sample data through GCN, the model can focus on the hydrogeological features that have a greater impact on the groundwater level among multiple categories by means of a spatial attention mechanism that assigns weights, improving the training accuracy of the model, and thus enabling more accurate prediction of the groundwater level.

[0120] In a possible implementation, the time feature extraction module includes a recurrent neural network and a temporal attention mechanism. A recurrent neural network is a neural network used to process time series data. The recurrent neural network maintains state information by forming internal feedback connections, thereby being able to capture the time-dependent relationship of the groundwater level at different time periods. For example, LSTM is a type of recurrent neural network. The temporal attention mechanism is used to assign different weights to the features of different time periods. During the process of predicting the groundwater level, the temporal attention mechanism can help identify which time period's feature data is the most critical for predicting the future groundwater level and accordingly increase the weights of these time periods.

[0121] Based on the structure of this time feature extraction module, the initial spatio-temporal feature vector can be obtained through processing by a recurrent neural network according to the spatial feature vector. The initial spatial feature vector is a preliminary feature representation calculated by the recurrent neural network based on the spatial feature vectors of different time periods.

[0122] The temporal attention mechanism generates different second weights for the features of different time periods in the initial spatio-temporal feature vector. The importance of the initial spatio-temporal feature vectors of different time periods for predicting the groundwater level is different. The temporal attention mechanism can assign corresponding second weights to the initial spatio-temporal feature data of different time periods, thereby measuring the importance of each time period for the groundwater level through the second weights. For example, since the dependence values of groundwater level prediction on time series of different lengths are different, larger second weight values can be assigned to adjacent time periods, while smaller second weight values can be assigned to time periods that are farther apart.

[0123] Then, based on the initial spatio-temporal feature vector and the second weights corresponding to different time periods in the initial spatio-temporal feature vector, the spatio-temporal feature vector is determined.

[0124] For example, referring to Figure 3 Formula 7, Formula 8, and Formula 9, for the input sequence X = [x1, x2,..., x T at time step (time period) t, the hidden state sequence of LSTM is H = [h1, h2,..., h T, the correlation score between each hidden state and the current hidden state is calculated through the ReLU activation function, and the attention β of each feature is generated through Softmax normalization (Equation 7). The hidden state is multiplied by the attention weight and weighted and summed to obtain the attention context vector (Equation 8). Finally, it is concatenated with the hidden state at the current moment (Equation 9) to obtain the final output result.

[0125] β = TA(H) = [β1, β2,..., f k 1×k (7)

[0126]

[0127] y T = Dense([h T , c T ) (9)

[0128] Thus, more accurate spatio-temporal feature vectors are generated through the recurrent neural network and the temporal attention mechanism. Specifically, during the model training process, more attention is paid to the feature data of the periods that are more important for groundwater level prediction. As a result, the groundwater level prediction model can not only capture the spatial characteristics of the groundwater level, but also deeply analyze the variation law in its time dimension, significantly improving the prediction accuracy.

[0129] See Figure 4 , which is a schematic flow chart of a method for training a groundwater level prediction model provided by an embodiment of the present application. First, the hydrogeological feature data collected from each monitoring point are collected and preprocessed. Then, an adjacency matrix is constructed to characterize the spatial correlation relationship between different monitoring points. The periods corresponding to the groundwater level labels and the feature data can be determined by means of a sliding window, so as to obtain a training sample set. By constructing a GCN-LSTM-Attention model, which includes GCN, LSTM, a temporal attention mechanism, and a spatial attention mechanism, the model is trained with the training sample set. Based on the model accuracy evaluation, the initial prediction model is continuously optimized. When the accuracy requirement is not met, the model parameters are adjusted continuously. When the accuracy requirement is met, the training is stopped to obtain a groundwater level prediction model with relatively high prediction accuracy. The groundwater level prediction model can predict the groundwater level for a certain future period or multiple future periods and output the prediction result.

[0130] Next, through Figure 5 and Figure 6 , a comprehensive and exemplary description of the training method provided by the embodiments of the present application is given, where Figure 5 is a schematic flow chart of a training technical architecture provided by an embodiment of the present application, Figure 6 is a schematic diagram of a training principle provided by an embodiment of the present application. ​

[0131] As shown Figure 5 in the figure, six steps are required to train the groundwater level prediction model.

[0132] Step 1: Collect dynamic data and static data through each monitoring point. The dynamic data includes groundwater level (used as the groundwater level label), rainfall, extraction volume, ecological water replenishment volume, etc., and the static data includes longitude and latitude, digital elevation value, rainfall infiltration coefficient, land use rate, etc. The initial data collected is preprocessed through means such as outlier processing, duplicate value deletion, missing value processing, and wavelet denoising.

[0133] Step 2: Construct an adjacency matrix based on the geographical locations of each monitoring point to characterize the spatial correlation relationship between different monitoring points.

[0134] Step 3: According to multi-category hydrological feature data and the time series data of groundwater level divided by a time sliding window, input features and output targets (groundwater level labels) are obtained, and a training sample set for training the model is constructed.

[0135] Step 4: Construct a GCN-LSTM-Attention model, input the training sample set into this model, and through GCN, spatial attention mechanism, LSTM, time attention mechanism, and fully connected layers, obtain the output result of the initial prediction model (predicted groundwater level).

[0136] Among them, as Figure 6 shown in the figure, the dynamic data, adjacency matrix, and static data corresponding to n monitoring points are respectively input into the model according to multiple time steps. The feature data of different time steps are extracted by GCN, and then strengthened by the spatial attention mechanism based on the first weight to obtain a spatial feature vector, which is then input into LSTM to capture the time series dependence relationship. The important time steps are strengthened by the time attention mechanism based on the second weight to obtain a spatio-temporal feature vector, so as to predict the current output result through the initial prediction model.

[0137] Step 5: Adjust the model parameters by adjusting hyperparameters, Dropout, L2 regularization, adaptive learning rate, etc. to meet the model accuracy requirements.

[0138] Step 6: Evaluate the model accuracy of the output result of the initial prediction model. If the accuracy requirements are met, stop training.

[0139] After training the groundwater level prediction model through the above steps, it is possible to, according to Step 7, design the scenario for predicting the current groundwater level by inputting rainfall, extraction volume, ecological replenishment volume, etc., then predict the groundwater level of one day through the groundwater level prediction model, and then add the groundwater level of this day to the input data, so as to realize the rolling prediction of the groundwater level for multiple days in the future.

[0140] See Figure 7 , Figure 7 Figure 7 shows a training device for a groundwater level prediction model provided by an embodiment of the present application. The device 700 includes a determination unit 701, a first processing unit 702, a second processing unit 703, a prediction unit 704, and an adjustment unit 705.

[0141] The determination unit 701 is configured to determine the groundwater level sample data of multiple monitoring points, where the multiple monitoring points are multiple monitoring points from the same area and having a spatial association. The groundwater level sample data includes a groundwater level label, the hydrogeological characteristic data of the monitoring point, and spatial characteristic data, and the spatial characteristic data is used to indicate the spatial association relationship between the multiple monitoring points.

[0142] The first processing unit 702 is configured to perform feature extraction on the groundwater level sample data of the multiple monitoring points through a spatial feature extraction module to obtain a spatial feature vector, and the spatial feature extraction module is used to capture the spatial association relationship between different monitoring points.

[0143] The second processing unit 703 is configured to process the spatial feature vector through a time feature extraction module to obtain a spatio-temporal feature vector, and the time feature extraction module is used to capture the change relationship of the groundwater level over time series.

[0144] The prediction unit 704 is configured to perform prediction through an initial prediction model according to the spatio-temporal feature vector to obtain the predicted groundwater level of each monitoring point.

[0145] The adjustment unit 705 is configured to adjust the model parameters of the initial prediction model according to the difference between the predicted groundwater level of each monitoring point and the groundwater level label, so as to obtain the groundwater level prediction model, so that the groundwater level prediction model can be used to predict the groundwater level of multiple monitoring points.

[0146] As can be seen from the above technical solution, for multiple monitoring points from the same region and with spatial correlation, the groundwater level sample data of the multiple monitoring points are determined according to the determination unit 701, where the groundwater level sample data includes groundwater level labels, hydrogeological characteristic data of the monitoring points, and spatial characteristic data. Since the hydrogeological characteristics of different monitoring points are correlated with each other in the spatial dimension, the groundwater level corresponding to each monitoring point is affected by the water level fluctuations at other locations. Therefore, by determining the spatial characteristic data indicating the spatial correlation relationship between the multiple monitoring points, a data source in the spatial dimension can be provided for training the model.

[0147] The first processing unit 702 extracts features from the groundwater level sample data of the multiple monitoring points through the spatial feature extraction module to obtain a spatial feature vector, thereby being able to capture the spatial correlation relationship between different monitoring points. The second processing unit 703 processes the spatial feature vector through the time feature extraction module to obtain a spatio-temporal feature vector, thereby being able to further capture the change relationship of the groundwater level over time on the basis of the spatial feature vector. The prediction unit 704 makes a prediction through the initial prediction model according to the spatio-temporal feature vector to obtain the predicted groundwater levels of each monitoring point. The adjustment unit 705 adjusts the model parameters of the initial prediction model according to the difference between the predicted groundwater levels of each monitoring point and the groundwater level labels to obtain a groundwater level prediction model for predicting the groundwater levels of the multiple monitoring points. Thus, by first determining the groundwater level sample data that can reflect the spatial correlation relationship between different monitoring points and capturing this correlation relationship through the spatial feature extraction module, during the process of training the model, not only can the future groundwater level be predicted based on the historical pattern presented by the groundwater level in the past period, but also the model can be adjusted so that the model can recognize the impact of the water level fluctuations at different locations with spatial correlation on the corresponding groundwater level of a certain monitoring point, thereby obtaining a groundwater level prediction model with the ability of spatial correlation analysis and improving the accuracy of the model in predicting the groundwater level.

[0148] As a possible implementation manner, the device further includes a rolling prediction unit for:

[0149] Obtain the historical groundwater level data of the multiple monitoring points, where the historical groundwater level data is used to indicate the groundwater level state information from the (i - n)-th period to the (i - 1)-th period in the past, n is a positive integer greater than 1 and less than i, i is a positive integer, and the (i - n)-th period to the (i - 1)-th period are multiple consecutive periods distributed in chronological order;

[0150] Predict the historical groundwater level data through the groundwater level prediction model to obtain the first groundwater level in the future i-th period;

[0151] Predict the historical groundwater level data and the first groundwater level through the groundwater level prediction model to obtain the second groundwater level at the (i + 1)-th future period.

[0152] As a possible implementation, the determining unit 701 is specifically configured to:

[0153] Obtain the geographical distribution information of the multiple monitoring points;

[0154] Determine the spatial association relationship between any two monitoring points in the area according to the geographical distribution information to obtain a plurality of spatial association relationships;

[0155] Construct an adjacency matrix of the multiple monitoring points according to the multiple spatial association relationships to obtain the spatial feature data.

[0156] As a possible implementation, the multiple monitoring points include a first monitoring point and a second monitoring point, and the determining unit 701 is specifically configured to:

[0157] Determine the spatial distance between the first monitoring point and the second monitoring point according to the geographical distribution information;

[0158] Determine the spatial association weight value between the first monitoring point and the second monitoring point according to the spatial distance. The spatial relationship weight value is used to identify the spatial association strength between the monitoring points. The smaller the spatial distance, the larger the spatial association weight value;

[0159] Determine the spatial association relationship between the first monitoring point and the second monitoring point according to the spatial association weight value;

[0160] Take any two monitoring points as the first monitoring point and the second monitoring point respectively to obtain the multiple spatial association relationships.

[0161] As a possible implementation, the spatial feature extraction module includes a graph convolutional neural network and a spatial attention mechanism. The first processing unit 702 is specifically configured to:

[0162] Extract features through the graph convolutional neural network according to the hydrogeological feature data of the monitoring points and the spatial feature data to obtain an initial spatial feature vector;

[0163] Generate different first weights for different categories of features in the initial spatial feature vector through the spatial attention mechanism;

[0164] Determine the spatial feature vector according to the initial spatial feature vector and the first weights corresponding to different categories of features in the initial spatial feature vector.

[0165] As a possible implementation manner, the time feature extraction module includes a recurrent neural network and a temporal attention mechanism, and the second processing unit 703 is specifically configured to:

[0166] Process the spatial feature vector through the recurrent neural network to obtain the initial spatio-temporal feature vector;

[0167] Generate different second weights for the features in different time periods of the initial spatio-temporal feature vector through the temporal attention mechanism;

[0168] Determine the spatio-temporal feature vector according to the initial spatio-temporal feature vector and the second weights respectively corresponding to different time periods in the initial spatio-temporal feature vector.

[0169] As a possible implementation manner, the hydrogeological feature data includes dynamic sample data and static sample data. The dynamic sample data is data that changes with time periods, including groundwater level, precipitation, extraction volume, and ecological replenishment volume. The static sample data is data that does not change with time periods, including longitude and latitude, digital elevation value, rainfall infiltration coefficient, and land use rate.

[0170] See Figure 8 , this application embodiment also provides a computer device, and the computer device includes a memory 801 and a processor 802:

[0171] The memory is used to store a computer program and transmit the computer program to the processor;

[0172] The processor is used to execute the method of the above method embodiment according to the computer program.

[0173] This application embodiment also provides a computer-readable storage medium, which is characterized in that the computer-readable storage medium is used to store a computer program, and the computer program is used to execute the method of the above method embodiment.

[0174] This application embodiment also provides a computer program product including a computer program. When it runs on a computer device, it causes the computer device to execute the method of the above method embodiment.

[0175] It should be noted that the various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other. For the system or device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0176] As used herein, the term "comprising" and its variations are open-ended, i.e., "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The relevant definitions of other terms will be given in the following description.

[0177] It should be understood that in this application, "at least one (item)" means one or more, and "a plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that there can be three relationships. For example, "A and / or B" can mean: only A exists, only B exists, and both A and B exist at the same time. Among them, A and B can be singular or plural. The character " / " generally means that the associated objects before and after are in an "or" relationship. "At least one (item) of the following" or its similar expression refers to any combination of these items, including any combination of a single item (item) or plural items (items). For example, at least one (item) of a, b, or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0178] It should also be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variation thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including the said element.

[0179] The steps of the method or algorithm described in connection with the embodiments disclosed herein can be implemented directly in hardware, a software module executed by a processor, or a combination of both. The software module can be placed in a random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0180] The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not intended to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A training method for a groundwater level prediction model, characterized in that: The method comprises: Determine groundwater level sample data of a plurality of monitoring points, the plurality of monitoring points being from the same region and having spatial association, the groundwater level sample data comprising groundwater level labels, hydrogeological characteristic data of the monitoring points, and spatial characteristic data, the spatial characteristic data being used to indicate the spatial association relationship between the plurality of monitoring points; Extracting features from the groundwater level sample data of the plurality of monitoring points by a spatial feature extraction module to obtain a spatial feature vector, wherein the spatial feature extraction module is used to capture the spatial correlation relationship between the different monitoring points; Processing the spatial feature vector through a time feature extraction module to obtain a spatiotemporal feature vector, wherein the time feature extraction module is used to capture the relationship between the groundwater level and the time series; According to the spatiotemporal characteristic vector, prediction is performed through an initial prediction model to obtain a predicted groundwater level at each of the monitoring points; According to the difference between the predicted groundwater level at each monitoring point and the groundwater level label, the model parameters of the initial prediction model are adjusted to obtain the groundwater level prediction model, so that the groundwater level prediction model can be used to predict the groundwater levels of multiple monitoring points.

2. The method according to claim 1, characterized in that: The method further comprises: Acquire the historical groundwater level data of the plurality of monitoring points, the historical groundwater level data being used to indicate the groundwater level status information from the inth time period to the i-1th time period in the past, wherein n is a positive integer greater than 1 and less than i, i is a positive integer, and the inth time period to the i-1th time period are a plurality of consecutive time periods distributed in chronological order; Predicting the historical groundwater level data using the groundwater level prediction model to obtain the first groundwater level in the i-th period in the future; The groundwater level historical data and the first groundwater level are predicted by the groundwater level prediction model to obtain a second groundwater level in the future i+1th period.

3. The method according to claim 1, characterized in that The spatial feature data is determined in the following manner: Obtaining geographical distribution information of the multiple monitoring points; Determine the spatial correlation relationship between any two monitoring points in the region according to the geographic distribution information to obtain multiple spatial correlation relationships; According to the multiple spatial association relationships, an adjacency matrix of the multiple monitoring points is constructed to obtain the spatial feature data.

4. The method according to claim 3, characterized in that The multiple monitoring points include a first monitoring point and a second monitoring point. The spatial correlation relationship between any two monitoring points in the region is determined according to the geographic distribution information to obtain multiple spatial correlation relationships, including: Determining a spatial distance between the first monitoring point and the second monitoring point according to the geographic distribution information; Determine a spatial association weight between the first monitoring point and the second monitoring point according to the spatial distance, wherein the spatial relationship weight is used to identify the spatial association strength between the monitoring points, and the smaller the spatial distance, the greater the spatial association weight; Determining a spatial association relationship between the first monitoring point and the second monitoring point according to the spatial association weight; Any two monitoring points are respectively used as the first monitoring point and the second monitoring point to obtain the multiple spatial association relationships.

5. The method according to claim 1, characterized in that The spatial feature extraction module includes a graph convolutional neural network and a spatial attention mechanism. The spatial feature extraction module extracts features from the groundwater level sample data of the multiple monitoring points to obtain a spatial feature vector, including: According to the hydrogeological characteristic data of the monitoring point and the spatial characteristic data, feature extraction is performed through the graph convolutional neural network to obtain an initial spatial feature vector; Generate different first weights for different categories of features in the initial spatial feature vector through the spatial attention mechanism; The spatial feature vector is determined according to the initial spatial feature vector and first weights respectively corresponding to features of different categories in the initial spatial feature vector.

6. The method according to claim 1, characterized in that The temporal feature extraction module includes a recurrent neural network and a temporal attention mechanism, and the temporal feature extraction module extracts features from the spatial feature vector to obtain a spatiotemporal feature vector, including: According to the spatial feature vector, the initial spatiotemporal feature vector is obtained by processing through the recurrent neural network; Generate different second weights for the features of different time periods in the initial spatiotemporal feature vector through the temporal attention mechanism; The spatiotemporal feature vector is determined according to the initial spatiotemporal feature vector and second weights respectively corresponding to different time periods in the initial spatiotemporal feature vector.

7. The method according to any one of claims 1 to 6, characterized in that: The hydrogeological characteristic data include dynamic sample data and static sample data. The dynamic sample data are data that change with time periods, including groundwater level, precipitation, extraction volume and ecological water replenishment volume. The static sample data are data that do not change with time periods, including longitude and latitude, digital elevation values, rainfall infiltration coefficient and land utilization rate.

8. A training device for a groundwater level prediction model, characterized in that: The device comprises a determination unit, a first processing unit, a second processing unit, a prediction unit and an adjustment unit; The determining unit is used to determine groundwater level sample data of a plurality of monitoring points, wherein the plurality of monitoring points are from the same region and are spatially associated, and the groundwater level sample data includes groundwater level labels, hydrogeological characteristic data of the monitoring points, and spatial characteristic data, wherein the spatial characteristic data is used to indicate the spatial association relationship between the plurality of monitoring points; The first processing unit is used to extract features from the groundwater level sample data of the plurality of monitoring points through a spatial feature extraction module to obtain a spatial feature vector, wherein the spatial feature extraction module is used to capture the spatial correlation relationship between different monitoring points; The second processing unit is used to process the spatial feature vector through a time feature extraction module to obtain a spatiotemporal feature vector, wherein the time feature extraction module is used to capture the relationship between the groundwater level and the time series; The prediction unit is used to perform prediction based on the spatiotemporal feature vector through an initial prediction model to obtain a predicted groundwater level at each monitoring point; The adjustment unit is used to adjust the model parameters of the initial prediction model according to the difference between the predicted groundwater level at each monitoring point and the groundwater level label to obtain the groundwater level prediction model so that the groundwater level prediction model can be used to predict the groundwater levels of multiple monitoring points.

9. A computer device, characterized in that: The computer device comprises a processor and a memory: The memory is used to store a computer program and transmit the computer program to the processor; The processor is configured to execute the method according to any one of claims 1 to 7 according to the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store a computer program, and the computer program is used to execute the method according to any one of claims 1 to 7.

Images