ConvLSTM-based regional groundwater supply capacity evaluation method

Abstract

The invention belongs to the technical field of water resource management, and discloses a ConvLSTM-based regional groundwater supply capacity assessment method, which comprises the steps of calculating a regional SPEI index, constructing a space-time cube by using the SPEI index, and determining space-time distribution of a regional drought condition; constructing a groundwater flow numerical model according to the hydrological parameters of the groundwater in the arid region, and identifying the water supply status of the groundwater; hydrological data of an evaluation area are acquired, weights of indexes and parameters are determined by using an entropy evaluation method, and a training set, a verification set and a test set are made; constructing a ConvLSTM model, and training the ConvLSTM model by using the training set; and inputting the test set into the trained ConvLSTM model, and evaluating the groundwater supply capacity. According to the invention, the model is established by using regional groundwater spatio-temporal data to accurately evaluate the groundwater supply capacity, so that groundwater resources can be utilized more effectively under drought conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of water resource management, and in particular relates to a method for evaluating regional groundwater supply capacity based on ConvLSTM. Background Art

[0002] Among the many freshwater resources on the earth, groundwater is undoubtedly the most abundant. Groundwater is an important source of drinking water in many regions. The stable groundwater supply capacity ensures people's basic living water needs, provides the necessary water for crop growth, and helps the normal operation and sustainable development of industrial enterprises, and protects the health and quality of life of local residents. Especially in some arid areas, groundwater supply capacity can be used as a reliable emergency water source to provide necessary water resources support for disaster-stricken areas and alleviate the adverse effects of disasters. However, groundwater level data are usually geographic spatiotemporal series data, and most current studies only consider spatial distance factors when predicting groundwater levels. In actual underground aquifers, groundwater levels are affected by factors such as the degree of rock formation development, topographic factors, and soil structure characteristics. Therefore, it is not enough to consider only spatial distance. It is necessary to consider the actual spatial connectivity characteristics, that is, the spatiotemporal characteristics of groundwater under certain hydrogeological conditions. In order to accurately evaluate the regional groundwater supply capacity under drought conditions, reduce excessive groundwater exploitation, and ensure its supply and demand balance under natural conditions, it is necessary to propose an effective evaluation method for regional water supply capacity under drought conditions. Summary of the invention

[0003] In view of the above defects or improvement needs of the prior art, the present invention aims to provide a new evaluation method for regional groundwater supply capacity under drought conditions by considering the temporal correlation of groundwater level data and the spatial connectivity between different observation wells.

[0004] To achieve the above object, the present invention provides a method for evaluating regional groundwater supply capacity based on ConvLSTM, the method comprising the following steps: S100: Calculate the regional SPEI index, construct a space-time cube using the SPEI index, and determine the spatial and temporal distribution of regional drought conditions; S200: Construct a numerical model of groundwater flow based on the hydrological parameters of groundwater in arid areas to identify the current status of groundwater supply; S300: Obtain hydrological data of the assessment area, use the entropy method to determine the weights of indicators and parameters, and create training sets, validation sets, and test sets; S400: Build a ConvLSTM model and train the ConvLSTM model using the training set; S500: Input the test set into the trained ConvLSTM model to evaluate the groundwater supply capacity.

[0005] Furthermore, the calculation of the regional SPEI index in S100 specifically includes: S111: Calculate potential evapotranspiration PET:

[0006] in, R n represents the net radiation, M Indicates the month, ET 0,PM represents the potential evapotranspiration, PM represents the monthly precipitation, S 2 means the wind speed 2 meters above the ground, v d represents the actual vapor pressure, v c represents the saturated vapor pressure, ▽ represents the slope of the change of temperature with the saturated water vapor pressure, T represents the average temperature, γ represents the hygrometer constant, G represents the soil heat flux density, U 2 is the wind speed at a height of 2 meters; S112: Calculate water gain or loss using PET values:

[0007] in, Dx is the precipitation-evapotranspiration difference, P x is the amount of precipitation at a fixed time each month; S113: Use Dx The SPEI index is calculated using the logarithmic logistic distribution and probability density function. The log-logistic distribution is as follows: The probability density function is:

[0008] in, α is the scale parameter, y is the basic parameter, β is the shape parameter; Calculate the value of SPEI: in, pis the probability of exceeding the water budget balance value, oh is the probability weighted distance, c0, c1, c c , d1, d2, and d3 are all constants.

[0009] Furthermore, the construction of a space-time cube using the SPEI index in S100 to determine the space-time distribution of drought conditions in each region specifically includes: S121: Three-dimensional visualization of the SPEI index in the region is performed based on the space-time cube model to obtain a space-time cube; S122: Perform K-means clustering on the time series set in the space-time cube and select the optimal number of clusters using the silhouette coefficient; S123: By slicing and dicing the data in the space-time cube, the distribution of spatial hot spots under different time slices and the changes of specific spatial areas in different time periods are analyzed; S124: Detect outliers that deviate from the surrounding data pattern through local anomaly analysis, and identify local clusters of outliers in the spatial data.

[0010] Furthermore, the step of constructing a groundwater flow numerical model according to the hydrological parameters of groundwater in arid areas and identifying the current status of groundwater supply in S200 specifically includes: S201: Calculate the permeability coefficient of the aquifer, The permeability of the unconfined aquifer is The permeability coefficient of the confined aquifer is The influence radius R is

[0011] In the above formula, Q is the pumping flow, H is the thickness of the unconfined aquifer, M is the thickness of the confined aquifer, r is the radius of the well, and S is the water level drawdown; S202: Based on the hydrological data and hydrogeological conditions of the assessment area, a two-dimensional groundwater flow numerical model is constructed using the linear infiltration principle of Darcy's law:

[0012] in, K x , K y For along x , yThe permeability coefficient in the coordinate axis direction, h For point ( x , y )exist t The water head value at the moment, h 0 is the initial hydraulic head of the aquifer, m s is the water content of the aquifer, W is the source-sink term, for C 2 The direction of the outer normal of the boundary, K n is the permeability coefficient in the direction of the boundary normal, q 1( x , y , t ) is the inflow rate per unit area on the second type boundary of the seepage zone, q 2 ( x , y , t ) is the outflow rate per unit area on the second type boundary of the seepage zone, C 2 is the second type of fixed flow boundary, C 3 is the second type of water-blocking boundary; S203: Use the measured hydrological data to verify the groundwater flow numerical model using the MODFLOW program; S204: Calculate the amount of groundwater in the assessment area based on the groundwater flow numerical model, and identify the current groundwater supply status in combination with the water demand of users in the area.

[0013] Furthermore, the method of using the entropy method to determine the weights of indicators and parameters in S300 specifically includes: S311: Perform Pearson correlation coefficient analysis on the collected hydrological data to screen the main control factors affecting water consumption. The formula is as follows:

[0014] in, x i and y i Respectively x and y The collection i elements, the number of elements in the set is n ; S312: Normalize the initial value and the positive index to

[0015] Negative indicators are normalized to

[0016] in, Xij ´For the first i The evaluation object is j The standardized value of the indicator, X ij is the original value; max ( X j )and minutes ( X j ) are the maximum and minimum values ​​respectively; S313: Calculate the j The first of the indicators i The proportion of indicators

[0017] The entropy value of each evaluation index is defined as

[0018] in, k =1 / ln m , m is the number of evaluation objects; if Y ij =0, then exclude the indicator j Evaluation object i ; S314: Calculate the redundancy of entropy:

[0019] S315: Calculate the weight of each evaluation index: .

[0020] Furthermore, the step of preparing the training set, the validation set and the test set in S300 specifically includes: S321: registering the collected sample label data to the corresponding active and passive fusion data to obtain sample data; S322: clipping the sample data into a sample data set of uniform size by using a regular grid clipping method; S323: Randomly divide the sample data set into a training set, a validation set, and a test set in a ratio of 6:2:2.

[0021] Furthermore, the ConvLSTM model structure in S400 is: in, X t Indicates at time t Input data; C t For the current moment t The cell state; H t For the current moment t The hidden state of it It is the input gate, which is used to determine how much new information can enter the cell state at the current moment; feet The forget gate is used to determine how much information about the cell state at the previous moment can be forgotten; oh The output gate is used to determine how much information about the cell state at the current moment can be output; W xi , W xf , W xo , W xc The input data are X t Convolution weight matrices to the input gate, forget gate, output gate, and cell state; W hi , W hf , W ho , W hc They are the hidden states of the previous moment. H t-1 Convolution weight matrices to the input gate, forget gate, output gate, and cell state; W ci , W cf , W co are the cell states at the previous moment. C t-1 Convolution weight matrices to the input gate, forget gate, and output gate; b i , b f , b o , b c They are the bias terms of input gate, forget gate, output gate and cell state respectively; " is the convolution operator; operator " "express Hadamard product; s represents a nonlinear activation function; tanh is the activation function.

[0022] Furthermore, constructing the ConvLSTM model in S400 also includes using a convolutional block attention module to enhance spatial features to improve the performance of the ConvLSTM model.

[0023] Furthermore, the use of the convolutional block attention module to enhance spatial features to improve the performance of the prediction model specifically includes: The channel attention module in the convolutional block attention module uses global average pooling to obtain global statistics of each channel and adopts two fully connected layers to learn channel weights; The spatial attention module in the convolutional block attention module uses max pooling and average pooling to extract detailed features from each spatial location; The two matrices obtained after the operation of the channel attention module and the spatial attention module are concatenated and processed through a convolutional layer and a sigmoid activation function to learn the weight of each spatial position; All the obtained weights are applied to each spatial position on the feature map.

[0024] Furthermore, the step of using the training set to train the ConvLSTM model in S400 specifically includes: S401: Pass the input data through 5 ConvLSTM layers to capture the water level variation pattern in local time and space; S402: Use the mean square error loss function to calculate the loss MSE between the output result and the true value: Among them, y i is the true value, is the predicted value, n is the number of samples; S403: Use the Adam optimization algorithm to repeatedly train the model and optimize the model parameters.

[0025] Compared with the prior art, the present invention can achieve the following beneficial effects: (1) The present invention uses the existing spatiotemporal data of regional groundwater under drought conditions based on the groundwater hydrological data of the assessment area to establish a model to accurately assess the water supply capacity of groundwater under drought conditions. By assessing the water supply capacity of regional groundwater under drought conditions, it is possible to more effectively utilize groundwater resources under drought or water shortage conditions, improve the utilization rate of groundwater resources, and reduce their losses.

[0026] (2) The present invention proposes a complete set of detailed groundwater assessment models for the entire process from groundwater recharge to discharge and use. Through the predictive evaluation of these models, the amount of groundwater resources under drought conditions can be calculated more effectively, which has obvious advantages in resisting drought disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A flowchart of a method for evaluating regional groundwater supply capacity based on ConvLSTM in a specific embodiment of the present invention; Figure 2 This is a structural diagram of the space-time cube model of a specific implementation method of the present invention.

[0028] Figure 3 This is the overall prediction process of the ConvLSTM model in the specific implementation manner of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0030] In order to accurately evaluate the water supply capacity of regional groundwater under drought conditions, this paper uses the spatiotemporal data of existing regional groundwater under drought conditions based on the hydrological indicators recorded in the observation wells and the hydrogeological characteristics of the monitoring site to establish a spatiotemporal data cube of groundwater and identify the supply and demand balance of regional groundwater under drought conditions. A convolutional long short-term memory (ConvLSTM) model is constructed to calculate and predict the water supply capacity of regional groundwater under drought conditions.

[0031] Reference Manual Attached Figure 1 As shown, the present invention provides a method for evaluating regional groundwater supply capacity based on ConvLSTM, and the method comprises the following steps: 1. Calculate the regional SPEI index and use the SPEI index to construct a space-time cube to determine the spatial and temporal distribution of regional drought conditions.

[0032] Collect regional hydrological data and calculate the Standardized Precipitation Evapotranspiration Index (SPEI). Use the SPEI index to assess the spatiotemporal evolution of regional drought and identify regional drought conditions. Analyze the characteristics and changes of drought conditions from the overall temporal and spatial scales. Use the spatiotemporal series model to visualize and analyze the multi-scale SPEI in time and space, and cluster the assessed areas using the K-means clustering method in spatiotemporal scales. Then use the pseudo F statistic to evaluate the optimal cluster number of multi-dimensional SPEI. Combine spatiotemporal hotspot analysis with local outliers to determine the development trend of drought conditions in each region.

[0033] 1.1 Calculate the Standardized Precipitation Evapotranspiration Index (SPEI) for a specific assessment area.

[0034] 1.1.1 Calculation of potential evapotranspiration, SPEI depends on the change in the difference between precipitation and potential evapotranspiration. There are many methods to calculate potential evapotranspiration (PET). Here we choose the Penman-Monteith calculation method to improve the accuracy of its results because it is more dependent on the atmospheric evapotranspiration demand. The calculation formula is as follows:

[0035] In this equation, R n represents the net radiation ( MJ / m 2 / d ), M Indicates the month, ET 0,PM represents the potential evapotranspiration ( mm ), PM represents the monthly precipitation, S 2 means the wind speed 2 meters above the ground ( m / s ), v d represents the actual vapor pressure ( kPa ), v c represents the saturated vapor pressure ( kPa ), ▽ represents the slope of the change of temperature with saturated water vapor pressure ( kPa / ℃), T represents the average temperature (°C), γ represents the hygrometer constant ( kPa / ℃), G represents the soil heat flux density ( MJ / m 2 / d ), U 2 is the wind speed at a height of 2 meters ( m / s ).

[0036] 1.1.2 Usage ET 0,PM The value is used to calculate the gain or loss of moisture in a fixed period of time. The calculation formula is as follows:

[0037] Dx is the precipitation-evapotranspiration difference, P x It is the amount of precipitation at a fixed time each month.

[0038] 1.1.3 Calculation D x The sum of .

[0039]

[0040]

[0041] For the u Year v Month starts k The difference in cumulative precipitation and evapotranspiration within a month, For the u -1st year m The difference in precipitation and evapotranspiration at the beginning of the month; For the u Year m The precipitation-evapotranspiration difference at the beginning of the month.

[0042] 1.1.4 Calculate the Standardized Precipitation Evapotranspiration Index (SPEI).

[0043] SPEI is a standardized Dx The three-parameter logistic distribution and probability density function are used to calculate the logistic distribution. The logistic distribution is as follows:

[0044] Dx The probability density function of is:

[0045] α is the scale parameter, y is the basic parameter, β is the shape parameter. Calculate the value of SPEI:

[0046]

[0047]

[0048] P is the probability of exceeding the water budget balance value, oh is the probability weighted distance. Among them,

[0049]

[0050] 1.1.5 Based on the SPEI value, the drought level standards are classified as shown in Table 1 below: Table 1 SPEI drought classification standards

[0051] 1.2 Use the SPEI drought index to assess the spatiotemporal evolution of regional drought and identify regional drought conditions.

[0052] Above, we mainly identified drought conditions based on the time dimension. Now, we will use the space-time cube model (which is mainly designed for multidimensional data analysis in geographic space, as shown in the attached figure). Figure 2 As shown in the figure, based on the concept of data cube (a representation of multidimensional data sets), by considering the relationship between time, space and other attributes, the SPEI index is used to integrate the information of the two dimensions of time and space for multidimensional data analysis, and the drought conditions of different observation points in the geographic space under the space-time cube model analysis are obtained, and drought monitoring at different time scales is realized.

[0053] 1.2.1 In the space-time cube model, a space-time cube is created by aggregating sample data onto space-time columns. The space-time data can be visualized and analyzed in three forms: time series, integrated space, and time pattern. The X-axis and Y-axis represent the spatial position of the time period, and the Z-axis represents the time. The starting time is at the bottom layer, and the top layer is the most recent time. Each cube represents the attribute value of the time period, and the value can be distinguished by setting different colors. The present invention mainly aggregates the SPEI values ​​by calculating the data of the observation area. The SPEI values ​​per unit time are aggregated into a cube, and multiple vertically arranged cubes are arranged and aggregated in time sequence to obtain a space-time column, so each observation area will form a space-time column. Therefore, by aggregating SPEI values ​​of different time scales, different types of drought conditions in the observation area can be obtained.

[0054] 1.2.2 Based on the previous step, the SPEI index in the study area is visualized in three dimensions using the space-time cube model. You can flexibly select multiple different time scales for visualization and analysis, such as short-term drought and long-term drought, to obtain a more accurate spatial and temporal distribution of drought conditions.

[0055] 1.3 Based on the space-time cube model, the time series set in the space-time cube is spatially represented by the K-means clustering algorithm, and the silhouette coefficient (used to measure the closeness between each data point and its cluster and other clusters) is used to select the optimal number of clusters and evaluate the clustering performance.

[0056] 1.3.1 First, Thiessen polygons were created based on the mutual spatial distances between observation wells and clustered using the K-means clustering algorithm.

[0057] 1.3.2 Then, further clustering is performed based on the hydrogeological indicator values ​​recorded at each monitoring point. The average profile value of the current iteration of each indicator is calculated within the predefined minimum and maximum cluster numbers.

[0058] 1.3.3 Finally, the indicator with the largest average silhouette value is selected as the optimal number of clusters for the classified data. In this way, we classify the observation wells with strong spatial connectivity into the same category, while the observation wells with weak connectivity are divided into different categories.

[0059] 1.3.4 Use the pseudo F statistic to select the number of clusters. The pseudo F statistic can reflect the ratio of the total deviation square sum (SST) to the intra-cluster square sum (SSE), that is, the ratio of the similarity within the group and the difference between the groups. The larger the value, the more effective the clustering is. The formula is as follows:

[0060] in

[0061]

[0062] n is the number of elements, n i For the i The number of elements in a cluster, n c is the total number of clusters, n v is the total number of variables in the cluster features, V ij k For the i The first j Elements k the value of the variable, for k The mean value of a variable, For the i In the cluster k The mean value of a variable.

[0063] 1.4 By slicing and dicing the data in the space-time cube, analyze the distribution of spatial hotspots under different time slices and the changes in specific spatial areas in different time periods.

[0064] 1.4.1 Detect clustering characteristics in the geographic space of the evaluated region based on spatiotemporal hotspot analysis, including outliers, hotspot areas, and spatiotemporal trends in the region.

[0065] 1.4.2 Analyze the spatiotemporal variation and geographic distribution of data to identify clustering patterns, trends, or anomalies. Use spatiotemporal hotspot analysis models to identify drought hotspots. Detect hotspots that have not experienced drought in the past, have milder droughts, or have experienced worsening droughts compared to a certain period of time.

[0066] 1.5 This is achieved by detecting outliers that deviate from the surrounding data patterns through local anomaly analysis, identifying local clusters of outliers in spatial data, and thus dividing the data into six categories: statistically significant high-value clusters, low-value clusters, high-value outliers mainly surrounded by low values, low-value outliers mainly surrounded by high values, meaningless values, and multiple types of values, in order to discover local anomalies, clusters, and abnormal patterns in the data.

[0067] 2. Construct a groundwater flow numerical model based on the hydrological parameters of groundwater in arid areas to identify the current status of groundwater supply.

[0068] By obtaining the hydrological parameters of groundwater in the identified arid area and constructing a groundwater flow numerical model, the current exploitable groundwater volume data in the assessment area can be obtained based on the model. The current groundwater supply status can be identified based on the water demand of users in the assessment area and the exploitable groundwater volume.

[0069] 2.1 Through the analysis of hydrological data combined with the current daily use of groundwater wells, it can be concluded that the main sources of groundwater recharge are rainfall runoff and infiltration and diffusion of river networks, followed by rainfall infiltration, irrigation infiltration, and lateral runoff recharge. The main sources of groundwater consumption are evaporation and artificial exploitation, among which artificial exploitation is the main form.

[0070] 2.2 Determine the permeability coefficient K.

[0071] 2.2.1 Using the Qiubuyi formula in the steady flow calculation formula, the formula for calculating the permeability coefficient K of the unconfined aquifer is:

[0072] The permeability coefficient K of the confined aquifer is

[0073] The influence radius R is

[0074] In the above formula, Q is the pumping flow rate ( m 3 / s ), H is the thickness of the unconfined aquifer ( m ), M is the thickness of the confined aquifer ( m ), r is the radius of the well ( m ), S The water level drops ( m ).

[0075] 2.2.2 Calculate the thickness of the aquifer H. The thickness of the aquifer is mainly determined by drawing a topographic elevation map of the assessment area based on the hydrogeological data in the assessment area, and then subtracting the data of the groundwater level depth map from it to obtain a rough map of the groundwater level elevation. The elevation contour map is then converted into an elevation grid map, and the aquifer thickness can be obtained by subtracting the elevation grid from the water level elevation map.

[0076] 2.3 Based on the hydrological data and hydrogeological conditions of the assessment area, a two-dimensional groundwater flow numerical model is constructed using the linear infiltration principle of Darcy's law. Since the output and input of the groundwater system in time and space are constantly changing, it is an unsteady flow. The two-dimensional unsteady flow model of the groundwater flow numerical model in the study area is:

[0077]

[0078]

[0079]

[0080] Where: K x , K y For along x , y Permeability coefficient in the coordinate axis direction ( m / d ); h For point ( x , y )exist t Water head value at the moment ( m ); h 0 is the initial hydraulic head of the aquifer ( m ); m s is the water content of the aquifer ( l / m );W is the source-sink term ( m 3 / d ); for C 2. The direction of the outer normal of the boundary; K n is the permeability coefficient in the direction of the boundary normal ( m / d ); q 1( x , y , t ) is the inflow rate per unit area on the second type boundary of the seepage zone ( m 3 / d ); q 2 ( x , y , t ) is the outflow rate per unit area on the second type boundary of the seepage zone ( m 3 / d ); C 2 is the second type of fixed flow boundary; C 3 is the second type of water-blocking boundary.

[0081] 2.4 A groundwater flow numerical model was constructed by analyzing the hydrogeological characteristics, and the model was validated using the MODFLOW program using the measured hydrological data.

[0082] 2.5 Based on the measured hydrological data of the assessment area, the initial underground water volume and the subsurface water infiltration volume are fitted to obtain the groundwater volume and water supply status of the assessment area.

[0083] 3. Obtain the hydrological data of the assessment area, use the entropy method to determine the weights of indicators and parameters, and create training sets, validation sets, and test sets.

[0084] 3.1 First, the entropy method is used to determine the weights of indicators and parameters based on the information provided by the observations.

[0085] 3.1.1 Perform Pearson correlation coefficient analysis on the collected hydrological historical data and select the main controlling factors affecting water consumption as the input of the model. The formula is as follows:

[0086] x i and y i Respectively x and y The collection i elements, the number of elements in the set is n .

[0087] 3.1.2 Normalize the initial value and the positive index to

[0088] Negative indicators are normalized to

[0089] Where: X ij ´For the first i The evaluation object is j Standardized values ​​on each indicator; X ij is the original value; max ( X j )and minutes ( X j ) are the maximum and minimum values ​​respectively.

[0090] 3.1.3 Calculation j The first of the indicators i The proportion of indicators

[0091] The entropy value of each evaluation index can be defined as

[0092] Where: k =1 / ln m , m is the number of evaluation objects. Y ij =0, then exclude the indicator j Evaluation object i .

[0093] 3.1.4 Calculate the redundancy of entropy, the formula is

[0094] 3.1.5 Calculate the weight of each evaluation indicator. The formula is: .

[0095] 3.2 Create training sets, validation sets, and test sets.

[0096] 3.2.1 Align the collected sample label data to the corresponding active and passive fusion data to obtain sample data.

[0097] 3.2.2 The sample data is cropped into a sample data set of size 25×25 by using the regular grid cropping method. The label of the sample data set is determined by the central pixel.

[0098] 3.2.3 The sample data set is randomly divided into training set, validation set and test set in a ratio of 6:2:2. During the training process, the hyperparameters of the model are adjusted through the validation set.

[0099] 4. Build a ConvLSTM model and train it using the training set.

[0100] Since groundwater level data is usually sequence data in time and space, the artificial neural network model (ANN) only considers the time distance factor when predicting groundwater level, but does not consider the actual spatial connectivity characteristics. Therefore, the present invention proposes to use the neural network prediction model combined with the connectivity characteristics of groundwater to predict the groundwater level in the assessment area under drought conditions. By evaluating the hydrogeological characteristics of the area, combining the advantages of convolutional neural network (CNN) and long short-term memory network (LSTM), a convolutional long short-term memory joint model (ConvLSTM) is constructed. ConvLSTM is a professional recurrent neural network. ConvLSTM is used to efficiently process high-dimensional data, and the weight sharing of data can be realized, thereby reducing the number of parameters of the model. The ConvLSTM unit can further improve the computational efficiency when processing different areas of data in parallel. The high-dimensional groundwater data is input into the ConvLSTM model, and the long sequence is divided into subsequences by the ConvLSTM model to process the one-dimensional time series data. The hydrological data of a fixed period of unit time is input to the ConvLSTM, so that it can learn the overall pattern of this period, so that the data can be reconstructed into segmented data for each time period, and at the same time, a more refined time pattern of each day can be captured. By learning the patterns of data changes over time, ConvLSTM extracts features on a time scale and makes accurate predictions about future states. This method has higher simulation accuracy than other methods. By simulating and predicting the groundwater level in the assessment area, the groundwater supply capacity of the area can be accurately assessed. The overall prediction process of the ConvLSTM model is shown in the attached figure. Figure 3 shown.

[0101] 4.1 Use Convolutional Block Attention Module (CBAM) to enhance spatial features to improve the performance of the prediction model.

[0102] CBAM consists of two attention modules: channel attention module and spatial attention module.

[0103] 4.1.1 The channel attention module in CBAM uses global average pooling to obtain global statistical information of each channel, and uses two fully connected layers to learn channel weights. During the propagation of input features, the values ​​in the channel attention module are first subjected to global average pooling, and the pooled values ​​are weighted normalized through a Sigmoid activation function. This process can be expressed by the following formula:

[0104] in, F represents the input feature value, represents F after the channel attention module, σ represents the activation function sigmoid , MLP represents a multi-layer perceptron, W 0 and W 1 represents the weight of the multi-layer perceptron, AvgPool and MaxPool Represent average pooling and maximum pooling respectively, and They are channel attention modules AvgPool ( F )and MaxPool ( F ) calculation results.

[0105] 4.1.2 Multiply the normalized weight value with the corresponding channel value to obtain the global statistical information of each channel. This process can be expressed by the following formula:

[0106] in, F ´ represents the feature value after being processed by the channel attention module, Indicates element-by-element multiplication. During the multiplication process, the attention value is diffused accordingly: the channel attention value is diffused along the spatial dimension.

[0107] 4.1.3 The spatial attention module in CBAM uses maximum pooling and average pooling to extract detailed features from each spatial position. In CBAM, multiple channels can be generated after convolution, and the spatial attention module applies maximum pooling and average pooling operations to each channel at each feature point. This process can be expressed by the following formula:

[0108] in, It means that F passes through the spatial attention module, represents a convolution operation using a 7×7 convolution kernel, and The spatial attention module AvgPool ( F )and MaxPool ( F ) calculation results.

[0109] 4.1.4 After the above channel attention module and spatial attention module are operated, two matrices will be obtained. The two matrices are connected and processed through the convolution layer and the Sigmoid activation function to learn the weight of each spatial position. This process can be expressed by the following formula: in, F " indicates the final output value, express F ´After the spatial attention module.

[0110] 4.1.5 Apply all the obtained weights to each spatial position on the feature map to enhance the spatial importance features. Using these attention modules can effectively improve the performance of the neural network model.

[0111] 4.2 Construct a ConvLSTM model with the following structure:

[0112]

[0113]

[0114]

[0115]

[0116] in, X t Indicates at time t Input data; C t For the current moment t The cell state; H t For the current moment t The hidden state of it It is the input gate, which is used to determine how much new information can enter the cell state at the current moment; feet The forget gate is used to determine how much information about the cell state at the previous moment can be forgotten; oh The output gate is used to determine how much information about the cell state at the current moment can be output; W xi , W xf , W xo , W xc The input data areX t Convolution weight matrices to the input gate, forget gate, output gate, and cell state; W hi , W hf , W ho , W hc They are the hidden states of the previous moment. H t-1 Convolution weight matrices to the input gate, forget gate, output gate, and cell state; W ci , W cf , W co are the cell states at the previous moment. C t-1 Convolution weight matrices to the input gate, forget gate, and output gate; b i , b f , b o , b c They are the bias terms of input gate, forget gate, output gate and cell state respectively; " is the convolution operator; operator " "express Hadamard product; s represents a nonlinear activation function; tanh is the activation function.

[0117] In the ConvLSTM prediction model of the present invention, there are 5 ConvLSTM layers, each of which has 8 filters of size (3,3) and stride 1. The same padding is used in each ConvLSTM layer to keep the spatial dimension of the spatiotemporal features extracted during the convolution process unchanged.

[0118] The raster data representing the groundwater supply capacity value of each grid point at a specific time point is converted into a 4D tensor with dimensions of width × height × time × feature.

[0119] After batch normalization of the input data, let it enter each input layer for operation and then perform reasonable activation.

[0120] After the first two batch normalization layers, 3D max pooling is performed with dimensions of (1, 2, 2) and (1, 3, 3), respectively.

[0121] After the first few ConvLSTM layers return sequences of the last K time steps, these sequences are fed to the next layer, and the last ConvLSTM layer returns the final long short-term features with a time dimension of 1.

[0122] These features are flattened and connected to a dense layer with 256 neurons, and the ReLU activation function is used for the dense layer to make the network converge faster.

[0123] A Dropout layer (regularization) with an intermediate probability of 0.5 is used between two dense layers. This layer can randomly ignore some neurons to avoid overfitting.

[0124] 4.3 Use the training set to train the model using the PyTorch library in Python.

[0125] 4.3.1 The input data is passed through 5 layers of ConvLSTM layers with a convolution kernel size of (3,3) to capture the water level change pattern in local time and space. The parameters can be adjusted appropriately. During the training process, the hyperparameters of the model are adjusted through the validation set.

[0126] 4.3.2 The loss function MSE is used to measure the difference between the predicted data and the actual data.

[0127] Select the mean square error (MSE) loss function to calculate the loss between the output result and the true value.

[0128]

[0129] where y i is the true value, is the predicted value and n is the number of samples.

[0130] 4.3.3 Use the optimization algorithm Adam to repeatedly train the model.

[0131] By repeatedly training the model through the above steps, its accuracy can be improved, thereby predicting the groundwater level.

[0132] 5. Input the test set into the trained ConvLSTM model to evaluate the groundwater supply capacity.

[0133] Based on historical experience, drought areas were selected as calculation and evaluation areas. According to the topography, they were divided into 25 water resource calculation units (areas). An observation well was set up in each calculation unit, and relevant hydrological data were obtained. The data were calculated and input into the ConvLSTM model to predict and evaluate the groundwater supply capacity.

[0134] The trained ConvLSTM model is used to predict groundwater levels, recharge changes, and the impact of groundwater on other water sources over a certain period of time, and accurately output the assessment results of groundwater supply capacity. These results can provide decision-making support for water resources management departments to ensure the sustainable use of groundwater resources and water supply security.

[0135] The above-mentioned embodiments only express the implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A method for evaluating regional groundwater supply capacity based on ConvLSTM, characterized in that: The method comprises the following steps: S100: Calculate the regional SPEI index, construct a space-time cube using the SPEI index, and determine the spatial and temporal distribution of regional drought conditions; S200: Construct a numerical model of groundwater flow based on the hydrological parameters of groundwater in arid areas to identify the current status of groundwater supply; S300: Obtain hydrological data of the assessment area, use the entropy method to determine the weights of indicators and parameters, and create training sets, validation sets, and test sets; S400: Build a ConvLSTM model and train the ConvLSTM model using the training set; S500: Input the test set into the trained ConvLSTM model to evaluate the groundwater supply capacity.

2. The method for evaluating regional groundwater supply capacity based on ConvLSTM according to claim 1, characterized in that: The calculation of the regional SPEI index in S100 specifically includes: S111: Calculate potential evapotranspiration PET: in, R n represents the net radiation, M Indicates the month, ET 0,PM represents the potential evapotranspiration, PM represents the monthly precipitation, S 2 means the wind speed 2 meters above the ground, v d represents the actual vapor pressure, v c represents the saturated vapor pressure, ▽ represents the slope of the change of temperature with the saturated water vapor pressure, T represents the average temperature, γ represents the hygrometer constant, G represents the soil heat flux density, U 2 is the wind speed at a height of 2 meters; S112: Calculate water gain or loss using PET values: in, Dx is the precipitation-evapotranspiration difference, P x is the amount of precipitation at a fixed time each month; S113: Use Dx The SPEI index is calculated using the logarithmic logistic distribution and probability density function. The log-logistic distribution is as follows: The probability density function is: in, α is the scale parameter, y is the basic parameter, β is the shape parameter; Calculate the value of SPEI: in, p is the probability of exceeding the water budget balance value, ω is the probability weighted distance, c0, c1, c c , d1, d2, and d3 are all constants.

3. The method for evaluating regional groundwater supply capacity based on ConvLSTM according to claim 2 is characterized in that: S100 uses the SPEI index to construct a space-time cube to determine the spatiotemporal distribution of drought conditions in each region, including: S121: Three-dimensional visualization of the SPEI index in the region is performed based on the space-time cube model to obtain a space-time cube; S122: Perform K-means clustering on the time series set in the space-time cube and select the optimal number of clusters using the silhouette coefficient; S123: By slicing and dicing the data in the space-time cube, the distribution of spatial hot spots under different time slices and the changes of specific spatial areas in different time periods are analyzed; S124: Detect outliers that deviate from the surrounding data pattern through local anomaly analysis, and identify local clusters of outliers in the spatial data.

4. The method for evaluating regional groundwater supply capacity based on ConvLSTM according to claim 1, characterized in that: S200 constructs a groundwater flow numerical model based on the hydrological parameters of groundwater in arid areas to identify the current status of groundwater supply, specifically including: S201: Calculate the permeability of the aquifer, The permeability of the unconfined aquifer is The permeability coefficient of the confined aquifer is The influence radius R is In the above formula, Q is the pumping flow, H is the thickness of the unconfined aquifer, M is the thickness of the confined aquifer, r is the radius of the well, and S is the water level drawdown; S202: Based on the hydrological data and hydrogeological conditions of the assessment area, a two-dimensional groundwater flow numerical model is constructed using the linear infiltration principle of Darcy's law: in, K x , K y For along x , y The permeability coefficient in the coordinate axis direction, h For point ( x , y )exist t The water head value at the moment, h 0 is the initial hydraulic head of the aquifer, μ s is the water content of the aquifer, W is the source-sink term, for Γ 2 The direction of the outer normal of the boundary, K n is the permeability coefficient in the direction of the boundary normal, q 1( x , y , t ) is the inflow rate per unit area on the second type boundary of the seepage zone, q 2 ( x , y , t ) is the outflow rate per unit area on the second type boundary of the seepage zone, Γ 2 is the second type of fixed flow boundary, Γ 3 is the second type of water-blocking boundary; S203: Use the measured hydrological data to verify the groundwater flow numerical model using the MODFLOW program; S204: Calculate the amount of groundwater in the assessment area based on the groundwater flow numerical model, and identify the current groundwater supply status in combination with the water demand of users in the area.

5. The method for evaluating regional groundwater supply capacity based on ConvLSTM according to claim 1, characterized in that: The method of using the entropy method to determine the weights of indicators and parameters in S300 specifically includes: S311: Perform Pearson correlation coefficient analysis on the collected hydrological data to screen the main control factors affecting water consumption. The formula is as follows: in, x i and y i Respectively x and y The collection i elements, the number of elements in the set is n ; S312: Normalize the initial value and the positive index to Negative indicators are normalized to in, X ij ´For the first i The evaluation object is j The standardized value of the indicator, X ij is the original value; max ( X j )and min ( X j ) are the maximum and minimum values ​​respectively; S313: Calculate the j The first of the indicators i The proportion of indicators The entropy value of each evaluation index is defined as in, k =1 / lnm , m is the number of evaluation objects; if Y ij =0, then exclude the indicator j Evaluation object i ; S314: Calculate the redundancy of entropy: S315: Calculate the weight of each evaluation index: 。 6. The method for evaluating regional groundwater supply capacity based on ConvLSTM according to claim 5, characterized in that: The preparation of the training set, the validation set and the test set in S300 specifically includes: S321: registering the collected sample label data to the corresponding active and passive fusion data to obtain sample data; S322: clipping the sample data into a sample data set of uniform size by using a regular grid clipping method; S323: Randomly divide the sample data set into a training set, a validation set, and a test set in a ratio of 6:2:

2.

7. The method for evaluating regional groundwater supply capacity based on ConvLSTM according to claim 1, characterized in that: The ConvLSTM model structure described in S400 is: in, X t Indicates at time t Input data; C t For the current moment t The cell state; H t For the current moment t The hidden state of it It is the input gate, which is used to determine how much new information can enter the cell state at the current moment; ft The forget gate is used to determine how much information about the cell state at the previous moment can be forgotten; ot The output gate is used to determine how much information about the cell state at the current moment can be output; W xi , W xf , W xo , W xc The input data are X t Convolution weight matrices to the input gate, forget gate, output gate, and cell state; W hi , W hf , W ho , W hc They are the hidden states of the previous moment. H t-1 Convolution weight matrices to the input gate, forget gate, output gate, and cell state; W ci , W cf , W co are the cell states at the previous moment. C t-1 Convolution weight matrices to the input gate, forget gate, and output gate; b i , b f , b o , b c They are the bias items of input gate, forget gate, output gate and cell state respectively;" " is the convolution operator; operator" "express Hadamard product; σ represents a nonlinear activation function; tanh is the activation function.

8. The method for evaluating regional groundwater supply capacity based on ConvLSTM according to claim 7, characterized in that: The construction of the ConvLSTM model described in S400 also includes using a convolutional block attention module to enhance spatial features to improve the performance of the ConvLSTM model.

9. The method for evaluating regional groundwater supply capacity based on ConvLSTM according to claim 8, characterized in that: The use of the convolutional block attention module to enhance spatial features to improve the performance of the prediction model specifically includes: The channel attention module in the convolutional block attention module uses global average pooling to obtain global statistics of each channel and adopts two fully connected layers to learn channel weights; The spatial attention module in the convolutional block attention module uses max pooling and average pooling to extract detailed features from each spatial location; The two matrices obtained after the operation of the channel attention module and the spatial attention module are concatenated and processed through a convolutional layer and a sigmoid activation function to learn the weight of each spatial position; All the obtained weights are applied to each spatial position on the feature map.

10. The method for evaluating regional groundwater supply capacity based on ConvLSTM according to claim 7, characterized in that: S400 uses the training set to train the ConvLSTM model, specifically including: S401: Pass the input data through 5 ConvLSTM layers to capture the water level variation pattern in local time and space; S402: Use the mean square error loss function to calculate the loss MSE between the output result and the true value: Among them, y i is the true value, is the predicted value, n is the number of samples; S403: Use the Adam optimization algorithm to repeatedly train the model and optimize the model parameters.

Images