Interpretable lake space water level simulation method and system considering space-time correlation

Abstract

The invention discloses an interpretable lake space water level simulation method and system considering time-space correlation, and belongs to the technical field of lake water level simulation. The method comprises the following steps: firstly, establishing a drainage basin graph structure and an adjacent matrix, constructing a lake space water level simulation model based on a graph space-time neural network, and segmenting a data set based on an equal-frequency box separation method; secondly, training a lake space water level simulation model, and preferably selecting a hyper-parameter combination; and finally, the influence of each inflow runoff on the lake space water level is quantified. The system includes a memory, a processor, and a computer program stored on the memory and operable on the processor. According to the invention, by capturing the time sequence correlation and the space correlation of each hydrological representative station in the drainage basin, the dynamic response process of the lake space water level to a plurality of runoffs entering the lake can be accurately simulated; on the basis of an interpretability method, time-space difference contribution of each inflow runoff to lake space water level fluctuation is analyzed, the limitation of black box of a traditional machine learning model is broken through, and accurate management and green development of a drainage basin can be achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of lake water level simulation, and relates to an interpretable lake space water level simulation method and system taking into account time and space correlation. Background Art

[0002] Lakes are an important component of global freshwater resources, playing an irreplaceable role in runoff regulation, flood prevention and disaster reduction, and ecological protection. Water level is a key indicator in lake water resource management. It not only reflects the hydrological situation inside the lake, but its fluctuation also profoundly affects many aspects such as river water and sediment transport, lake water quality conditions, and habitats. However, in recent years, the large-scale construction of upstream reservoirs has significantly changed the hydrological situation of lakes. The water level has dropped significantly during the impoundment period, causing the dry season to be advanced and extended, and even leading to a series of ecological problems such as deterioration of lake water quality and degradation of wetland ecological functions. In order to clarify the impact of upstream reservoirs on lake hydrology and ecology and scientifically guide reservoir scheduling, it is particularly important to accurately simulate the spatial water level response of lakes and clarify the response mechanism of lakes to water from different directions. However, under the joint influence of many rivers from different directions and with different runoff levels, the spatiotemporal distribution of lake water levels is highly complex and dynamic, which brings huge challenges to accurately simulating the spatial water level fluctuations of lakes.

[0003] The process model based on the physical equations for fine modeling of the watershed provides the possibility to accurately and comprehensively reflect the response of the lake's spatial water level to the river runoff. By constructing a detailed grid, the process model can simulate the spatiotemporal dynamic response of the lake's flooding morphology, spatial water level, water flow direction and other characteristics. For example, Chinese invention patent 202210071934.9 discloses a method for deep coupling of one-dimensional and two-dimensional hydrodynamic models of rivers and lakes, which divides the study area into several one-dimensional and two-dimensional partitions and connects them with reduced dimensions, and derives one-dimensional cross-sectional flow data and plane velocity field data step by step. However, although the process model can provide a deep understanding of the response of the lake's spatial water level to various runoffs, the model construction relies on complex terrain parameters with high spatiotemporal resolution, and its iterative solution requires high computational costs. This technical threshold not only restricts the regional applicability promotion of the model, but also makes it difficult to effectively couple with multi-objective optimization models, and cannot be applied to the optimization research of future reservoir scheduling rules.

[0004] With the development and innovation of computer technology, data-driven models have gradually become a new trend in basin hydrological modeling due to their extremely high computational efficiency and powerful learning ability for nonlinear correlations of high-dimensional data. In particular, deep learning models represented by LSTM and GRU have achieved remarkable results in time series data prediction and have been widely used. For example, Chinese invention patent 202410494656.7 discloses a lake water level conformal prediction method that integrates Copula function and deep learning, performs two-stage correction of cumulative distribution function and Copula function on the prediction results of LSTM model, considers the correlation between multiple random variables and the dependency between prediction time steps, and predicts lake water level based on the calibrated LSTM model. Chinese invention patent 202410730759.9 discloses a method for predicting saltwater tide upstream in estuary areas using machine learning algorithms, selects salinity sequences and exogenous driving factors as input, and inputs them into the gated recurrent unit GRU model to predict the salinity values ​​of saltwater tide upstream in estuary areas during the dry season at different forecast periods. However, these models focus on the extraction and prediction of time series features, and have obvious deficiencies when processing data with spatial correlation. For lakes, water level fluctuations are not only affected by the hydraulic exchange between rivers and lakes, but also by the intricate hydraulic connections within them, with close spatial correlations and mutual influences between different regions. Therefore, effectively capturing the spatial correlation between the internal runoff of the lake and that of the external runoff is crucial to improving the accuracy of the spatial water level simulation of the lake.

[0005] In addition to the challenges in model construction, deep learning models also face the problem of limited interpretability in analyzing the runoff response of lakes. Although they have shown extraordinary potential in simulation accuracy, they are often still regarded as "black boxes" and have difficulty explaining the contribution of various inflows, including reservoir discharge, to the hydrological fluctuations of the temporal and spatial heterogeneity of the lake, and cannot provide in-depth guidance on the response time and degree of the lake to reservoir regulation. Its opaque internal decision-making logic limits a deeper understanding of the runoff response mechanism and the optimization and guidance of basin regulation in the future. Summary of the invention

[0006] In view of the problems existing in the prior art, the present invention provides an interpretable lake spatial water level simulation method and system considering spatiotemporal correlation. On the basis of the traditional consideration of the temporal correlation of hydrological characteristics, the present invention takes into account the hydraulic impact of various inflows on the lake and the mutual correlation of the internal spatial water level of the lake, and fully integrates the temporal correlation and spatial correlation of the hydrological characteristics of the basin based on the graph spatiotemporal neural network, so as to accurately simulate the differentiated water level response process at multiple stations in different lake areas in the lake. In addition, based on the interpretability method, the present invention analyzes the response mechanism of the lake spatial water level to multiple inflows, quantifies the contribution of different inflows to the water level fluctuations of each station in the lake, and solves the problem of limited interpretability widely existing in existing deep learning models. The present invention combines advanced machine learning technology with an interpretable analysis framework, which can not only simulate the spatial water level of the lake efficiently and accurately, provide strong technical support for lake hydrological modeling, but also analyze the temporal and spatial differences of the contributions of various inflows, including reservoir discharge, to the spatial water level fluctuations of the lake, and provide a profound mechanism understanding and scientific guidance for basin water resources scheduling management and sustainable development.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for simulating lake water levels in an interpretable manner considering spatiotemporal correlations comprises the following steps:

[0009] Step 1, establish the watershed graph structure and adjacency matrix;

[0010] Step 2: construct a lake spatial water level simulation model based on graph spatiotemporal neural network;

[0011] Step 3, segment the data set based on the equal frequency binning method;

[0012] Step 4, training the lake spatial water level simulation model and optimizing the hyperparameter combination;

[0013] Step 5: quantify the impact of each inflow on the lake's spatial water level.

[0014] Further, step 1 is as follows:

[0015] Step 1.1, based on the river network near the lake, determine the lake's inflow and outflow channels, as well as the lake's internal divisions and flow directions. Select at least one hydrological representative station in each inflow river and lake division, and the selected hydrological representative station needs to have long-term, continuous, and consistent historical observation data. Based on differences in geographical location and hydrological characteristics, the hydrological representative stations include two types: lake runoff stations and lake water level stations.

[0016] Step 1.2, generalize the watershed graph structure.

[0017] The watershed graph structure consists of a node set and an edge set. The nodes are used to represent the hydrological representative stations selected in step 1.1, including the river runoff nodes representing the runoff stations into the lake. and lake level nodes representing lake level stations The edges represent the hydraulic connections between hydrological representative stations, including the edges outside the lake, from the river runoff node to the lake water level node. , characterizing the impact of runoff into the lake on the lake water level; and the edges that exist inside the lake and point from the lake water level node to the lake water level node , characterizing the correlation between the water levels inside the lake. Both types of edges are directed edges, and their directions are consistent with the direction of the water flow.

[0018] Step 1.3, set up the adjacency matrix.

[0019] The information update and transmission of the watershed graph structure is as shown in formula (1). The adjacency matrix A is represented by N, where N is the number of nodes in the watershed graph structure, including P river runoff nodes and Q lake water level nodes. If and only if there is an edge from node p to node q, the element .

[0020] (1)

[0021] Further, step 2 is as follows:

[0022] The input of the lake spatial water level simulation model is the runoff characteristics, including the runoff sequence of all runoff into the lake in the past and the current H time periods, and the output is the water level characteristics of all lake water level stations in the current period. As shown in the following formula,

[0023] (2)

[0024] in, is the runoff characteristic at time t, , P represents the number of river runoff nodes. is the water level characteristic at time t, , Q represents the number of lake level nodes. H is the number of time periods for input runoff characteristics, is the mapping function.

[0025] There is a significant temporal and spatial correlation between the input runoff characteristics and the output water level characteristics. Specifically, there is a long-term temporal dependence in the runoff sequence, and this dependence continues to act on the water level fluctuation at time t; at the same time, the runoff characteristics of the river entering the lake and water level characteristics of the lake's internal zones Together they constitute a complete graph structure at time t, and the spatial correlation between the two is also an important factor affecting the fluctuation of the lake's spatial water level. To this end, the present invention integrates three spatiotemporal correlation learning modules and constructs a lake spatial water level simulation model based on a graph spatiotemporal neural network. Specifically: First, a long short-term memory network module, abbreviated as LSTM module, is introduced to mine the temporal correlation of the runoff sequence, and the runoff sequence is aggregated and reduced in dimension from the time dimension. Secondly, a graph convolutional neural network module, abbreviated as GCN module, is constructed. Based on the spatial correlation between the runoff into the lake and the water level of the adjacent lake, the spatial water level of the lake is preliminarily simulated, and the initial eigenvalues ​​at time t are assigned to the lake water level nodes. Finally, the residual graph convolutional neural network module, abbreviated as ResGCN module, is used. On the basis of considering the impact of the runoff into the lake, the spatial correlation between the lake water level nodes is further considered to accurately simulate the spatial water level of the lake.

[0026] Step 2.1: Capture the temporal association and reduce the dimension of the runoff sequence based on the LSTM module. The LSTM module uses the runoff sequence of all P runoff stations entering the lake in the basin for the past and current H time periods. First, the LSTM unit of the LSTM module captures the temporal correlation in the runoff sequence of each river entering the lake in the basin, extracts and updates the runoff characteristics of each runoff station entering the lake at each time step, and outputs the hidden state of each time period Then, based on the fully connected layer of the LSTM module, the hidden states of each runoff station in H periods are transformed into Linear combination reduces the dimension of the lake runoff sequence from the time dimension, and performs nonlinear transformation through the activation function, and finally outputs the aggregated runoff characteristics of each lake runoff station .

[0027] Step 2.2: Preliminary calculation of the lake spatial water level based on the GCN module. The GCN module uses the aggregated runoff characteristics of each runoff station into the lake output by the LSTM module As input, we study the hydraulic impact of each runoff station on its neighboring lake water level station, and output the initial estimated water level characteristics of Q lake water level stations in the basin. , providing input for the subsequent ResGCN module to accurately estimate the lake spatial water level. In the GCN module, the spatial associations considered only include the edges from the river runoff node to the lake water level node, as described in step 1.2. The information transfer within the GCN module can be characterized as:

[0028] (3)

[0029] (4)

[0030] Among them, the GCN module contains one or several GCN layers. Represents the input features of the lth GCN layer; It represents the output features of the lth GCN layer and is also the input feature of the l+1th GCN layer; A is the adjacency matrix described in step 1.3, and I is the corresponding unit matrix. and are the adjacency matrix and degree matrix with self-loops superimposed on them, Used for normalization of the adjacency matrix. is the learnable weight matrix. Represents the sigmoid activation function.

[0031] Step 2.3, accurately simulate the lake spatial water level based on the ResGCN module. The ResGCN module first concatenates the runoff characteristics of the river entering the lake at time t in formula (1) The initial estimated water level characteristics of each lake water level station output by the GCN module , get all node features at time t , and use it as the input of the ResGCN module. Then, based on the ResGCN module, the lake spatial water level is accurately simulated by considering the impact of lake runoff on the water level of neighboring lakes and the mutual influence between the lake spatial water levels. In the ResGCN module, the spatial associations considered include the edges outside the lake as described in step 1.2, which point from the river runoff node to the lake water level node. , and also covers the edges that exist inside the lake and point from the lake water level node to the lake water level node The information transfer of the ResGCN module is shown in the following formula: first, the graph convolution operation is performed, and then the graph convolution result is combined with the input of the ResGCN module The sum is taken as the output of the ResGCN module.

[0032] (5)

[0033] Further, step 3 is as follows:

[0034] Step 3.1, data equal frequency binning. First, the historical observation data of each hydrological representative station described in step 1.1 are sorted, and the data outliers and missing values ​​are interpolated. Then, the processed historical observation data are normalized to the maximum and minimum to obtain a data set. Then, the data set is sorted with the mean water level of all lake water level stations in each period as an indicator and the hydrological characteristics of all hydrological representative stations in each period as units, and the data set is evenly divided into M bins according to the sorting, and the number of time periods containing data in each bin is the same.

[0035] Step 3.2, data set division. Divide the data set into training set, validation set and test set. Randomly extract 20% of the data samples from each bin interval to form an independent test set. And use the K-fold cross-validation method to randomly divide the remaining 80% of the non-test set samples in each bin into K subsets with the same number of time periods. Finally, combine the k-th subset of each bin into the k-th fold validation set, and combine the remaining subsets into the k-th fold training set. Where k=1,2,…,K.

[0036] Further, step 4 is as follows:

[0037] Step 4.1, hyperparameter setting. The lake spatial water level simulation model contains multiple hyperparameters, including the number of LSTM unit layers, the number of hidden nodes of LSTM units, batch size, learning rate, and weight decay rate. Multiple candidate values ​​are assigned to each hyperparameter, and all candidate values ​​of each hyperparameter are combined to construct a hyperparameter combination set.

[0038] Step 4.2, optimize the hyperparameter combination of the lake spatial water level simulation model. Select the optimization method and loss function, and use K-fold training sets to train the lake spatial water level simulation model under each hyperparameter combination to obtain the optimal model for each hyperparameter combination. Input K validation sets into the optimal model of each hyperparameter combination, select the accuracy evaluation index, and calculate the average accuracy of the validation set of the optimal model. Subsequently, use all non-test set samples to retrain the lake spatial water level simulation model under each hyperparameter combination, and input the test set into the optimal model of each hyperparameter combination obtained by retraining, and use the accuracy evaluation index to calculate the accuracy of the test set. Finally, the hyperparameter combination selected when both the average accuracy of the validation set and the accuracy of the test set are optimal is the preferred hyperparameter combination.

[0039] Further, step 5 is as follows:

[0040] Step 5.1, calculate the marginal contribution of runoff into the lake to the spatial water level fluctuation of the lake.

[0041] Select interpretation samples from the data set described in step 3.1, and call the SHAP algorithm to calculate the marginal contribution of each inflow into the lake, including reservoir discharge, to the spatial water level of the lake, which is represented by a SHAP value matrix with the dimension of the number of inflow stations × the number of lake water level stations.

[0042] Step 5.2, analyze the time lag of the impact of runoff into the lake on the spatial water level of the lake.

[0043] Taking the lake runoff station p and the lake water level station q as an example, the marginal contribution of the lake runoff station p to the lake water level station q under different lags is extracted from the SHAP value matrix. The lag of the lake runoff corresponding to the two or three largest SHAP values ​​is the lag of the impact of the lake runoff p on the lake water level station q.

[0044] Step 5.3, quantify the impact of runoff into the lake on the spatial water level of the lake.

[0045] The marginal contribution of each runoff into the lake at each lag time to the designated lake water level station q is normalized and percentage-ized to obtain the period runoff contribution rate; the period runoff contribution rates of the same runoff into the lake with different lag times are summed to obtain the total contribution rate of each runoff into the lake to the lake water level station q; the total contribution rates of each runoff into the lake are sorted from large to small, and a larger total contribution rate indicates that the corresponding runoff into the lake has a stronger impact on the lake water level station q.

[0046] A computer system comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the steps of the above method are implemented when the processor executes the computer program.

[0047] The present invention has the following beneficial effects:

[0048] The present invention is based on the LSTM module with superior performance in temporal association extraction and the GCN and ResGCN modules that can effectively capture the spatial association of the watershed, and constructs a lake spatial water level simulation model based on a graph spatiotemporal neural network. The model effectively learns the complex hydraulic impact of each inflow on the lake spatial water level by capturing the temporal association and spatial association of each hydrological representative station in the watershed, and accurately simulates the dynamic response process of the lake spatial water level to many inflows including upstream reservoir discharge. At the same time, based on the interpretability method, the present invention analyzes the response time of the lake spatial water level to each inflow, quantifies the degree of influence of different inflows on the lake water level fluctuations, breaks through the limitations of the "black box" of traditional machine learning models, and provides model support and scientific basis for lake water resources scheduling and management, which is helpful to achieve precise management and green development of watersheds. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A flow chart of an interpretable lake spatial water level simulation method considering spatiotemporal correlation provided by the present invention;

[0050] Figure 2 It is a schematic diagram of the structure in the embodiment. DETAILED DESCRIPTION

[0051] The present invention is further described below in conjunction with specific embodiments.

[0052] The present invention selects the Dongting Lake Basin to develop an interpretable lake spatial water level simulation method that considers temporal and spatial correlation. The specific implementation method is described in detail in combination with the technical scheme and the accompanying drawings, which specifically includes the following steps:

[0053] Step 1: Establish the watershed graph structure and adjacency matrix.

[0054] Step 1.1, based on the river network near the lake, determine the lake's inflow and outflow channels, as well as the lake's internal divisions and flow directions. Select at least one hydrological representative station in each inflow river and lake division, and the selected hydrological representative station needs to have long-term, continuous, and consistent historical observation data. Based on differences in geographical location and hydrological characteristics, the hydrological representative stations include two types: lake runoff stations and lake water level stations.

[0055] In this embodiment, Dongting Lake includes five inflows into the lake, namely the diversion of the Yangtze River mainstream in the Sankou area, and the confluence of the Lishui River, Yuanjiang River, Zishui River and Xiangjiang River in the Sishui area. The Three Gorges Reservoir, Shimen, Taoyuan, Taojiang and Xiangtan stations are selected in turn as the hydrological representative stations of the Yangtze River mainstream, Lishui River, Yuanjiang River, Zishui River and Xiangjiang River, which are also the inflow stations for deducing the lake spatial water level.

[0056] Dongting Lake is divided into three lake areas, namely West Dongting Lake, South Dongting Lake and East Dongting Lake according to the direction of water flow. Its representative hydrological stations are Nanzui and Xiaohezui stations, Yangliutan and Yingtian stations, Lujiao and Chenglingji stations, which are the lake water level stations for deducing the lake spatial water level.

[0057] Step 1.2, generalize the watershed graph structure.

[0058] The watershed graph structure consists of a node set and an edge set. The nodes are used to represent the hydrological representative stations selected in step 1.1, including the river runoff nodes representing the runoff stations into the lake. and lake level nodes representing lake level stations The edges represent the hydraulic connections between hydrological representative stations, including the edges outside the lake, from the river runoff node to the lake water level node. , characterizing the impact of runoff into the lake on the lake water level; and the edges that exist inside the lake and point from the lake water level node to the lake water level node , characterizing the correlation between the water levels inside the lake. Both types of edges are directed edges, and their directions are consistent with the direction of the water flow. The graph structure of this embodiment is as follows Figure 2 shown.

[0059] Step 1.3, set the adjacency matrix. The information update and transmission of the watershed graph structure is as shown in formula (1) with the dimension The adjacency matrix A is represented by N, where N is the number of nodes in the watershed graph structure, including P river runoff nodes and Q lake water level nodes. If and only if there is an edge from node p to node q, the element In this embodiment, including P=5 river runoff nodes and Q=6 lake water level nodes, the dimension of the adjacency matrix A of the graph structure is 11×11.

[0060] (1)

[0061] Step 2: construct a lake spatial water level simulation model based on graph spatiotemporal neural network;

[0062] A lake spatial water level simulation model is constructed. The input of the lake spatial water level simulation model is the runoff characteristics, including the runoff sequence of all runoff into the lake in the past and the current H time periods, and the output is the water level characteristics of all lake water level stations in the current period. As shown in the following formula,

[0063] (2)

[0064] in, is the runoff characteristic at time t, , P represents the number of river runoff nodes, and its value is 5 in the embodiment. is the water level characteristic at time t, , Q represents the number of lake water level nodes, and in the embodiment, the value is 6. H is the number of time periods for inputting runoff characteristics, and in the embodiment, the value is 50. is the mapping function.

[0065] There is a significant temporal and spatial correlation between the input runoff characteristics and the output water level characteristics. Specifically, there is a long-term temporal dependence in the runoff sequence, and this dependence continues to act on the water level fluctuation at time t; at the same time, the runoff characteristics of the river entering the lake and water level characteristics of the lake's internal zones Together they constitute a complete graph structure at time t, and the spatial correlation between the two is also an important factor affecting the fluctuation of the lake's spatial water level. To this end, the present invention integrates three spatiotemporal correlation learning modules and constructs a lake spatial water level simulation model based on a graph spatiotemporal neural network. Specifically: First, a long short-term memory network module, abbreviated as LSTM module, is introduced to mine the temporal correlation of the runoff sequence, and the runoff sequence is aggregated and reduced in dimension from the time dimension. Secondly, a graph convolutional neural network module, abbreviated as GCN module, is constructed. Based on the spatial correlation between the runoff into the lake and the water level of the adjacent lake, the spatial water level of the lake is preliminarily simulated, and the initial eigenvalues ​​at time t are assigned to the lake water level nodes. Finally, the residual graph convolutional neural network module, abbreviated as ResGCN module, is used. On the basis of considering the impact of the runoff into the lake, the spatial correlation between the lake water level nodes is further considered to accurately simulate the spatial water level of the lake.

[0066] Step 2.1: Capture the temporal association and reduce the dimension of the runoff sequence based on the LSTM module. The LSTM module uses the runoff sequence of all P runoff stations entering the lake in the basin for the past and current H time periods. First, the LSTM unit of the LSTM module captures the temporal correlation in the runoff sequence of each river entering the lake in the basin, extracts and updates the runoff characteristics of each runoff station entering the lake at each time step, and outputs the hidden state of each time period Then, based on the fully connected layer of the LSTM module, the hidden states of each runoff station in H periods are transformed into Linear combination reduces the dimension of the lake runoff sequence from the time dimension, and performs nonlinear transformation through the activation function, and finally outputs the aggregated runoff characteristics of each lake runoff station .

[0067] Step 2.2: Preliminary calculation of the lake spatial water level based on the GCN module. The GCN module uses the aggregated runoff characteristics of each runoff station into the lake output by the LSTM module As input, we study the hydraulic impact of each runoff station on its neighboring lake water level station, and output the initial estimated water level characteristics of Q lake water level stations in the basin. , providing input for the subsequent ResGCN module to accurately estimate the lake spatial water level. In the GCN module, the spatial associations considered only include the edges from the river runoff node to the lake water level node, as described in step 1.2. The information transfer within the GCN module can be characterized as:

[0068] (3)

[0069] (4)

[0070] Among them, the GCN module contains one or several GCN layers. Represents the input features of the lth GCN layer; It represents the output features of the lth GCN layer and is also the input feature of the l+1th GCN layer; A is the adjacency matrix described in step 1.3, and I is the corresponding unit matrix. and are the adjacency matrix and degree matrix with self-loops superimposed on them, Used for normalization of the adjacency matrix. is the learnable weight matrix. Represents the sigmoid activation function. The GCN module constructed in this embodiment includes 1 GCN layer.

[0071] Step 2.3, accurately simulate the lake spatial water level based on the ResGCN module. The ResGCN module first concatenates the runoff characteristics of the river entering the lake at time t in formula (1) The initial estimated water level characteristics of each lake water level station output by the GCN module , get all node features at time t , and use it as the input of the ResGCN module. Then, based on the ResGCN module, the lake spatial water level is accurately simulated by considering the impact of lake runoff on the water level of neighboring lakes and the mutual influence between the lake spatial water levels. In the ResGCN module, the spatial associations considered include the edges outside the lake as described in step 1.2, which point from the river runoff node to the lake water level node. , and also covers the edges that exist inside the lake and point from the lake water level node to the lake water level node The information transfer of the ResGCN module is shown in the following formula: first, the graph convolution operation is performed, and then the graph convolution result is combined with the input of the ResGCN module The sum is taken as the output of the ResGCN module.

[0072] (5)

[0073] Step 3, segment the data set based on the equal frequency binning method;

[0074] Step 3.1, data is binned with equal frequency. First, the historical observation data of each hydrological representative station described in step 1.1 are sorted, and the data outliers and missing values ​​are interpolated, and then the processed historical observation data are normalized to the maximum and minimum to obtain a data set. Among them, the historical observation data selected in this embodiment are the daily-scale flow or water level data of all hydrological representative stations in the basin from 2009 to 2023, and the runoff characteristic of the Three Gorges Reservoir is the downstream flow. Subsequently, the data set is sorted with the mean water level of all lake water level stations in each time period as an indicator and the hydrological characteristics of all hydrological representative stations in each time period as units, and the data set is evenly divided into M bins according to the sorting, and the number of time periods containing data in each bin is the same. In this embodiment, the number of bins is set to 30.

[0075] Step 3.2, data set division. The data set is divided into a training set, a validation set, and a test set. Randomly extract 20% of the data samples from each bin interval to form an independent test set. The embodiment adopts a 5-fold cross-validation method to randomly divide the remaining 80% of the non-test set samples in each bin into 5 subsets with the same number of time periods. Finally, the k-th subset of each bin is combined into the k-th fold validation set, and the remaining subsets are combined into the k-th fold training set. Wherein, k=1,2,3,4,5.

[0076] Step 4, training the lake spatial water level simulation model and optimizing the hyperparameter combination;

[0077] Step 4.1, hyperparameter setting. The lake spatial water level simulation model contains multiple hyperparameters, including the number of layers of LSTM units, the number of hidden nodes of LSTM units, batch size, learning rate and weight decay rate. Assign multiple candidate values ​​to each hyperparameter, and combine all candidate values ​​of each hyperparameter to construct a hyperparameter combination set. In this embodiment, the candidate value range of the number of layers of LSTM units is [1,2,3], the candidate value range of the number of hidden nodes of LSTM units is [8,16,32], the candidate value range of batch size is [32,64,128], the candidate value range of learning rate is [0.01,0.05,0.001,0.005], and the candidate value range of weight decay rate is [0.01,0.001,0.0001], so the hyperparameter combination set includes 3×3×3×4×3=324 hyperparameter combinations.

[0078] Step 4.2, the hyperparameter combination of the lake spatial water level simulation model is optimized. The embodiment selects the Adam optimization method and the loss function of minimizing the mean square error, and uses a 5-fold training set to train the lake spatial water level simulation model under each hyperparameter combination to obtain the optimal model for each hyperparameter combination. The five validation sets are input into the optimal model of each hyperparameter combination, and the accuracy evaluation index is selected to calculate the average accuracy of the validation set of the optimal model. Subsequently, all non-test set samples are used to retrain the lake spatial water level simulation model under each hyperparameter combination, and the test set is input into the optimal model of each hyperparameter combination obtained by retraining, and the accuracy evaluation index is used to calculate the test set accuracy. Finally, the hyperparameter combination selected when both the validation set average accuracy and the test set accuracy reach the optimal level is the preferred hyperparameter combination.

[0079] In this embodiment, when the number of layers of the LSTM unit is 1, the number of hidden nodes of the LSTM unit is 8, the batch size is 64, the learning rate is 0.001, and the weight decay rate is 0.0001, the average accuracy of the validation set and the accuracy of the test set of the lake spatial water level simulation model are both optimal. At this time, the Pearson correlation coefficient and root mean square error of the validation set are 0.984 and 0.44m respectively, and the Pearson correlation coefficient and root mean square error of the test set are 0.985 and 0.42m respectively, both of which are global optimal.

[0080] Step 5: quantify the impact of each inflow on the lake's spatial water level.

[0081] Step 5.1, calculate the marginal contribution of runoff into the lake to the spatial water level fluctuation of the lake.

[0082] Select an explanation sample from the data set described in step 3.1, call the SHAP algorithm to calculate the marginal contribution of each runoff into the lake, including reservoir discharge, to the spatial water level of the lake, and represent it as a SHAP value matrix with a dimension of the number of runoff stations into the lake × the number of lake water level stations. This embodiment selects the independent test set described in step 3.2 as the explanation sample, and calculates the marginal contribution of the runoff series of the Three Gorges Reservoir, Shimen, Taoyuan, Taojiang, and Xiangtan stations to the water level fluctuations of Nanzui, Xiaohezui, Yangliutan, Yingtian, Lujiao, and Chenglingji stations, that is, a SHAP value matrix with a dimension of 5×6.

[0083] Step 5.2, analyze the time lag of the impact of runoff into the lake on the spatial water level of the lake.

[0084] Taking the lake runoff station p and the lake water level station q as an example, the marginal contribution of the lake runoff station p to the lake water level station q under different lags is extracted from the SHAP value matrix. The lag of the lake runoff corresponding to the two or three largest SHAP values ​​is the lag of the impact of the lake runoff p on the lake water level station q.

[0085] This embodiment analyzes the time lag of the impact of the Three Gorges Reservoir discharge on the water levels of Nanzui, Yangliutan and Chenglingji stations. For Nanzui station, the three marginal contributions of the Three Gorges Reservoir discharge with the largest values ​​are 0.0103, 0.0097 and 0.0093, and the corresponding time lags are 0 days, 2 days and 1 day, respectively; for Yangliutan station, the three marginal contributions of the Three Gorges Reservoir discharge with the largest values ​​are 0.0069, 0.0068 and 0.0060, and the corresponding time lags are 3 days, 2 days and 1 day, respectively; for Chenglingji station, the three marginal contributions of the Three Gorges Reservoir discharge with the largest values ​​are 0.0119, 0.0114 and 0.0107, and the corresponding time lags are 3 days, 2 days and 4 days, respectively. Therefore, the time lags of the impact of the Three Gorges Reservoir discharge on the water levels of Nanzui, Yangliutan and Chenglingji stations are 0-2 days, 1-3 days and 2-4 days, respectively.

[0086] Step 5.3, quantify the impact of runoff into the lake on the spatial water level of the lake.

[0087] The marginal contribution of each runoff into the lake at each lag time to the designated lake water level station q is normalized and percentage-ized to obtain the period runoff contribution rate; the period runoff contribution rates of the same runoff into the lake with different lag times are summed to obtain the total contribution rate of each runoff into the lake to the lake water level station q; the total contribution rates of each runoff into the lake are sorted from large to small, and a larger total contribution rate indicates that the corresponding runoff into the lake has a stronger impact on the lake water level station q.

[0088] This embodiment quantifies the influence of the runoff sequence of the Three Gorges Reservoir, Shimen, Taoyuan, Taojiang, and Xiangtan stations on the water levels of Nanzui and Yingtian stations. Among them, the total contribution rates of the Three Gorges Reservoir, Shimen, Taoyuan, Taojiang, and Xiangtan stations to the Nanzui station are 40%, 17%, 25%, 10%, and 8%, respectively. Therefore, the influence of each runoff into the lake on the water level fluctuation of the Nanzui station is from strong to weak: the main stream of the Yangtze River, Yuanjiang River, Lishui River, Zishui River, and Xiangjiang River; the total contribution rates of the Three Gorges Reservoir, Shimen, Taoyuan, Taojiang, and Xiangtan stations to the Yingtian station are 40%, 15%, 18%, 10%, and 17%, respectively. Therefore, the influence of each runoff into the lake on the water level fluctuation of the Yingtian station is from strong to weak: the main stream of the Yangtze River, Yuanjiang River, Xiangjiang River, Lishui River, and Zishui River.

[0089] The above-described embodiments merely express the implementation methods of the present invention, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for those skilled in the art, 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.

Claims

1. An interpretable lake spatial water level simulation method considering spatiotemporal correlation, characterized in that: The following steps are involved: Step 1, establish the watershed graph structure and adjacency matrix; Step 2: construct a lake spatial water level simulation model based on graph spatiotemporal neural network; The long short-term memory network module, graph convolutional neural network module, and residual graph convolutional neural network module are integrated to construct a lake spatial water level simulation model based on graph spatiotemporal neural network. First, the long short-term memory network module, abbreviated as LSTM module, is introduced to mine the temporal correlation of runoff sequences, aggregate and reduce the dimension of runoff sequences from the time dimension. Secondly, a graph convolutional neural network module, abbreviated as GCN module, is constructed to preliminarily simulate the lake spatial water level based on the spatial correlation between the runoff into the lake and the water level of the adjacent lakes, and the initial eigenvalue of the lake water level node at time t is assigned. Finally, the residual graph convolutional neural network module, abbreviated as ResGCN module, is used to further consider the spatial correlation between the lake water level nodes on the basis of considering the impact of the runoff into the lake, and accurately simulate the lake spatial water level. The input of the lake spatial water level simulation model is the runoff characteristics, including the runoff sequence of all runoff into the lake in the past and the current H time periods, and the output is the water level characteristics of all lake water level stations in the current period; as shown in the following formula, (2) in, is the runoff characteristic at time t, , P represents the number of river runoff nodes; is the water level characteristic at time t, , Q represents the number of lake water level nodes; H is the number of time periods for input runoff characteristics, is the mapping function; Step 3, segment the data set based on the equal frequency binning method; Step 4, training the lake spatial water level simulation model and optimizing the hyperparameter combination; Step 5: quantify the impact of each runoff into the lake on the lake's spatial water level, specifically: Step 5.1, calculate the marginal contribution of runoff into the lake to the spatial water level fluctuation of the lake; Step 5.2, analyzing the time lag of the impact of runoff into the lake on the water level of the lake space; Step 5.3, quantify the impact of runoff into the lake on the spatial water level of the lake.

2. The interpretable lake spatial water level simulation method considering spatiotemporal correlation according to claim 1 is characterized in that: The step 1 is specifically as follows: Step 1.1, based on the river network near the lake, determine the inflow and outflow channels of the lake, as well as the internal divisions and flow directions of the lake; select at least one hydrological representative station in each inflow channel and lake division, and the selected hydrological representative station needs to have long-term, continuous and consistent historical observation data; according to the differences in geographical location and hydrological characteristics, the hydrological representative stations include two types: inflow runoff stations and lake water level stations; Step 1.2, generalize the watershed graph structure; The watershed graph structure consists of a node set and an edge set; the nodes are used to represent the hydrological representative stations selected in step 1.1, including the river runoff nodes representing the runoff stations into the lake. and lake level nodes representing lake level stations ; Edges represent hydraulic connections between hydrological representative stations, including edges outside the lake that point from river runoff nodes to lake water level nodes. , characterizing the impact of runoff into the lake on the lake water level; and the edges that exist inside the lake and point from the lake water level node to the lake water level node , representing the correlation between the water levels inside the lake; both types of edges are directed edges, and their directions are consistent with the direction of water flow; Step 1.3, set the adjacency matrix; The information update and transmission of the watershed graph structure is as shown in formula (1). The adjacency matrix A is represented by N, where N is the number of nodes in the watershed graph structure, including P river runoff nodes and Q lake water level nodes; an element ; (1) 。 3. The interpretable lake spatial water level simulation method considering spatiotemporal correlation according to claim 2 is characterized in that: The step 2 is specifically as follows: Step 2.1, based on the LSTM module, the temporal correlation of the runoff sequence is captured and the dimension is reduced; The LSTM module uses the past and current runoff sequences of all P runoff stations entering the lake in the basin for a total of H time periods. As input; firstly, the LSTM unit of the LSTM module captures the temporal correlation in the runoff sequence of each river entering the lake in the basin, extracts and updates the runoff characteristics of each runoff station entering the lake at each time step, and outputs the hidden state of each time period ; Then, based on the fully connected layer of the LSTM module, the hidden states of each runoff station in the H time periods are transformed through the weight matrix. Linear combination reduces the dimension of the lake runoff sequence from the time dimension, and performs nonlinear transformation through the activation function, and finally outputs the aggregated runoff characteristics of each lake runoff station ; Step 2.2, preliminarily calculate the lake spatial water level based on the GCN module; The GCN module uses the aggregated runoff characteristics of each runoff station into the lake output by the LSTM module As input, we study the hydraulic impact of each runoff station on its neighboring lake water level station, and output the initial estimated water level characteristics of Q lake water level stations in the basin. , providing input for the subsequent ResGCN module to accurately estimate the lake spatial water level; Step 2.3, accurately simulate the lake spatial water level based on the ResGCN module; The ResGCN module first concatenates the runoff characteristics of the river entering the lake at time t in formula (1): The initial estimated water level characteristics of each lake water level station output by the GCN module , get all node features at time t , and use it as the input of the ResGCN module; then based on the ResGCN module, the lake spatial water level is accurately simulated by considering the impact of lake runoff on the water level of adjacent lakes and the mutual influence between the lake spatial water levels. and output it.

4. The interpretable lake spatial water level simulation method considering spatiotemporal correlation according to claim 3 is characterized in that: In step 2.2, the information transfer within the GCN module is characterized as follows: (3) (4) Among them, the GCN module contains one or several GCN layers. Represents the input features of the lth GCN layer; represents the output features of the lth GCN layer, and is also the input feature of the l+1th GCN layer; A is the adjacency matrix described in step 1.3, and I is the corresponding unit matrix; and are the adjacency matrix and degree matrix with self-loops superimposed on them, Used for normalization of adjacency matrix; is a learnable weight matrix; Represents the sigmoid activation function.

5. The interpretable lake spatial water level simulation method considering spatiotemporal correlation according to claim 3 is characterized in that: In step 2.3, the information transfer of the ResGCN module is shown in formula (5). First, the graph convolution operation is performed, and then the graph convolution result is combined with the input of the ResGCN module. Sum as the output of the ResGCN module; (5) 。 6. The interpretable lake spatial water level simulation method considering spatiotemporal correlation according to claim 3 is characterized in that: The step 3 is as follows: Step 3.1, data is divided into equal frequency bins; Arrangement step 1.1: The historical observation data of each hydrological representative station are interpolated for data outliers and missing values, and the processed historical observation data are normalized to the maximum and minimum to obtain a data set; Subsequently, the data set is sorted using the mean water level of all lake water level stations in each period as an indicator and the hydrological characteristics of all hydrological representative stations in each period as units, and the data set is evenly divided into M bins according to the sorting, with the same number of periods of data in each bin; Step 3.2, divide the dataset into training set, validation set and test set.

7. The interpretable lake spatial water level simulation method considering spatiotemporal correlation according to claim 6 is characterized in that: The step 3.2 is specifically as follows: randomly extract 20% of the data samples from each bin interval to form an independent test set; and use the K-fold cross-validation method to randomly divide the remaining 80% of the non-test set samples in each bin into K subsets with the same number of time periods; finally, combine the kth subset of each bin into the kth fold validation set, and combine the remaining subsets into the kth fold training set, where k=1,2,…,K.

8. The interpretable lake spatial water level simulation method considering spatiotemporal correlation according to claim 6 is characterized in that: The step 4 is specifically as follows: Step 4.1, hyperparameter setting; The lake spatial water level simulation model contains multiple hyperparameters, including the number of LSTM unit layers, the number of hidden nodes of LSTM units, batch size, learning rate, and weight decay rate; multiple candidate values ​​are assigned to each hyperparameter, and all candidate values ​​of each hyperparameter are combined to construct a hyperparameter combination set; Step 4.2, optimizing the hyperparameter combination of the lake spatial water level simulation model; Select the optimization method and loss function, use K-fold training sets to train the lake spatial water level simulation model under each hyperparameter combination, and obtain the optimal model for each hyperparameter combination; input K validation sets into the optimal model of each hyperparameter combination, select the accuracy evaluation index, and calculate the average accuracy of the validation set of the optimal model; Subsequently, all non-test set samples were used to retrain the lake spatial water level simulation model under each hyperparameter combination, and the test set was input into the optimal model of each hyperparameter combination obtained by retraining, and the test set accuracy was calculated using the accuracy evaluation index; finally, the hyperparameter combination selected when both the average accuracy of the validation set and the accuracy of the test set reached the optimal level is the preferred hyperparameter combination.

9. The interpretable lake spatial water level simulation method considering spatiotemporal correlation according to claim 8 is characterized in that: The step 5 is specifically as follows: Step 5.1, calculate the marginal contribution of runoff into the lake to the spatial water level fluctuation of the lake; Select interpretation samples from the data set, call the SHAP algorithm to calculate the marginal contribution of each inflow into the lake, including reservoir discharge, to the lake's spatial water level, represented by a SHAP value matrix with the dimension of the number of inflow stations × the number of lake water level stations; Step 5.2, analyzing the time lag of the impact of runoff into the lake on the water level of the lake space; Take the lake runoff station p and the lake water level station q as an example, and extract the marginal contribution of the lake runoff station p to the lake water level station q under different lag times from the SHAP value matrix. The lag time of the lake runoff corresponding to the largest 2 or 3 SHAP values ​​is the lag time of the impact of the lake runoff p on the lake water level station q. Step 5.3, quantify the impact of runoff into the lake on the lake's spatial water level; The marginal contribution of each runoff into the lake at each lag time to the designated lake water level station q is normalized and percentage-ized to obtain the period runoff contribution rate; the period runoff contribution rates of the same runoff into the lake with different lag times are summed to obtain the total contribution rate of each runoff into the lake to the lake water level station q; the total contribution rates of each runoff into the lake are sorted from large to small, and a larger total contribution rate indicates that the corresponding runoff into the lake has a stronger impact on the lake water level station q.

10. A computer system for implementing the interpretable lake spatial water level simulation method considering spatiotemporal correlation as described in any one of claims 1 to 9, characterized in that: The computer system includes a memory, a processor, and a computer program stored in the memory and executable on the processor.

Images