Hydrological forecasting method and system integrating confluence time-space relationship

Abstract

The invention belongs to the technical field of hydrology and computers, and discloses a hydrology forecasting method and system integrated with confluence time-space relationship, and the method comprises the steps: calculating the runoff production results of a plurality of sub-basins in a preset region according to a runoff production module of a Xinanjiang model; converting the runoff production results of the plurality of sub-basins into a plurality of first runoff production results; and inputting the plurality of first runoff production results into a convolutional long-short-term memory neural network model for confluence calculation to obtain a confluence result of the preset area, and performing hydrological forecasting according to the confluence result. According to the method, the interaction of flood waves among the sub-basins is accurately captured through attention calculation, the runoff generation sequence is modeled through feature mapping in combination with the convolutional long-short-term memory neural network, the confluence space-time relationship is fully considered, and the capability of capturing the space-time dependency relationship in the basins by the model is improved. The method is applied to small and medium-sized drainage basins, can ensure accurate hydrological forecasting even if data is limited, has high adaptability and wide application prospects, and provides powerful support for drainage basin management and flood control decision making.

Description

Technical Field

[0001] The invention discloses a hydrological forecasting method and system integrating the spatial and temporal relationship of confluence, belonging to the technical field of hydrology and computer science. Background Art

[0002] The Xin'anjiang model is a lumped hydrological model, and a decentralized model can also be selected according to the size of the basin area. It is widely used in humid and semi-humid areas, especially in humid seasons. The model divides the entire basin into multiple unit sub-basins, and calculates the runoff of each unit basin to obtain the outlet flow process of the unit basin. On this basis, river flood routing is performed to obtain the flow process line of the basin outlet.

[0003] However, the confluence calculation module of the Xin'anjiang model is insufficient in expressing the nonlinear laws of the basin. Especially in small and medium-sized basins, due to the small basin area, there are obvious nonlinear effects in the flood process. The traditional Xin'anjiang model uses the linear reservoir method and Muskingum method for confluence calculation. These methods cannot consider the interaction of flood waves between sub-basins, and it is difficult to accurately express the complex nonlinear confluence process of small and medium-sized basins, resulting in large errors. In addition, the parameters of the confluence process are highly sensitive and are greatly affected by human engineering and natural climate factors. Especially when the observation data is limited, it is difficult for the model to maintain accuracy. Summary of the invention

[0004] The purpose of the present invention is to provide a hydrological forecasting method and system that integrates the temporal and spatial relationship of confluence to solve the problems existing in the current technology. Specifically, the traditional Xin'anjiang model uses the linear reservoir method and the Muskingum method to calculate the confluence. These methods fail to fully consider the interaction of flood waves between sub-basins and are difficult to accurately describe the complex nonlinear confluence process in small and medium-sized basins. In other words, in the confluence calculation, these methods fail to accurately take into account the temporal and spatial relationship, resulting in a large confluence error. To achieve the above purpose, the present invention proposes a hydrological forecasting method that integrates the temporal and spatial relationship of confluence, including:

[0005] Step 1: Calculate the runoff results of multiple sub-basins in the preset area according to the runoff module of the Xin'anjiang model;

[0006] Step 2, converting the runoff generation results of the plurality of sub-watersheds into a plurality of first runoff generation results respectively;

[0007] Step 3: Input the plurality of the first runoff generation results into a convolutional long short-term memory neural network model to perform runoff calculation, obtain the runoff results of the preset area, and perform hydrological forecasting based on the runoff results.

[0008] Preferably, step 2 specifically includes:

[0009] Step 2.1, extracting the features of the runoff results of the plurality of sub-basins to generate a plurality of feature tensors;

[0010] Step 2.2: perform structural adjustment on each feature tensor, and record the flow generation result corresponding to each feature tensor after the structural adjustment as the first flow generation result.

[0011] Preferably, before the feature map conversion, the method further includes:

[0012] Attention calculation is performed on the flow generation results of the multiple sub-basins.

[0013] Preferably, performing attention calculation on the flow generation results of the plurality of sub-watersheds specifically includes:

[0014] Performing convolution operations on the runoff generation results of the plurality of sub-basins respectively, and generating a plurality of groups of feature sequences according to the runoff generation results of the plurality of sub-basins;

[0015] Performing nonlinear spatiotemporal attention scoring on each group of feature sequences to obtain an attention score for each group of feature sequences;

[0016] A weighted calculation is performed on each group of sequence features according to the attention score to generate multiple groups of weighted feature representations.

[0017] Preferably, nonlinear spatiotemporal attention scoring is performed on each group of feature sequences, specifically including:

[0018] Determine all data points in each group of the feature sequences according to a preset time and a preset spatial position;

[0019] The spatiotemporal attention score of the data point is determined according to its characteristic value, and the set of the spatiotemporal attention scores of all data points is recorded as the nonlinear spatiotemporal attention score of each group of the feature sequences.

[0020] Preferably, weighted calculation is performed on each group of sequence features according to the attention score to generate multiple groups of weighted feature representations, specifically including:

[0021] Obtaining the sum of the attention scores of all spatial positions at each time point in each group of sequence features, and recording it as the time point attention score;

[0022] Determine the probability value of each data point according to the spatiotemporal attention score of each data point and the time point attention score of the time point at which it is located;

[0023] A weighted calculation is performed on each group of sequence features according to the probability value to generate multiple groups of weighted feature representations.

[0024] Preferably, step 2 specifically includes:

[0025] Extracting feature tensors of multiple groups of weighted feature representations, and performing structural adjustment on each of the feature tensors;

[0026] The structurally adjusted feature tensor is mapped and transformed according to the parity of the time index, and the feature representation corresponding to the feature tensor after the mapping transformation is recorded as the first flow generation result.

[0027] Preferably, the convolutional long short-term memory neural network model is optimized and trained by an error direction propagation algorithm.

[0028] Preferably, the loss function of the convolutional long short-term memory neural network model is mean square error;

[0029] The convolutional long short-term memory neural network model saves the best model through the early stopping method.

[0030] A hydrological forecasting system integrating the spatial and temporal relationship of runoff, including the runoff generation module of the Xin'anjiang model and the convolutional long short-term memory neural network model;

[0031] The runoff generation module of the Xin'anjiang model is used to calculate the runoff generation results of multiple sub-basins in a preset area;

[0032] The convolutional long short-term memory neural network model couples the attention mechanism with the feature mapping method, and is used to convert the runoff generation results of the multiple sub-basins into multiple first runoff generation results, and perform confluence calculation based on the multiple first runoff generation results.

[0033] Compared with the prior art, the present invention has the following beneficial effects: the present invention effectively captures the interaction of flood waves between sub-basins through attention computing technology, and realizes accurate learning of the correlation between any two sub-basins; through the feature mapping method, the spatiotemporal relationship of the confluence process is fully considered, and the flow generation sequence is modeled in combination with the convolutional long short-term memory neural network, which further enhances the performance in capturing the spatiotemporal dependency relationship within the basin. When applied to small and medium-sized basins, the forecasting method of the present invention shows extremely high stability and reliability, so that accurate hydrological forecasts can be guaranteed even when data is limited, greatly improving the efficiency of basin management and the accuracy of flood control decisions. In addition, it has strong adaptability and broad application prospects, and can be flexibly applied to various basin environments, bringing significant social and economic benefits to the fields of water resources management, flood prevention and control, and ecological protection, and is of great significance for protecting people’s lives and property and promoting sustainable economic and social development. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A schematic diagram of a hydrological forecast process in an embodiment of the present invention;

[0035] Figure 2 This is a schematic diagram of the attention mechanism structure in an embodiment of the present invention;

[0036] Figure 3 Schematic diagram of the unit structure of the convolutional long short-term memory neural network in an embodiment of the invention. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] Embodiment 1:

[0039] See also Figure 1-3 As shown, the purpose of this embodiment is to provide a hydrological forecasting method integrating the temporal and spatial relationship of confluence, including:

[0040] Step 1: Calculate the runoff results of multiple sub-basins in the preset area according to the runoff module of the Xin'anjiang model;

[0041] Specifically, the measured data of ground hydrology, rainfall and evaporation stations in the preset area are obtained, including daily rainfall, daily runoff and water surface evaporation; the acquired data are preprocessed to ensure data quality and consistency, and the digital elevation data (DEM) is used to extract the watershed system and divide the preset area into sub-basins; the three-layer evaporation model of the Xin'anjiang model is used to calculate the runoff generation in the preset area, and the runoff generation results of each sub-basin in the preset area are obtained.

[0042] Step 2, converting the runoff generation results of the multiple sub-basins into multiple first runoff generation results respectively;

[0043] Furthermore, step 2 specifically includes: step 2.1, extracting the characteristics of the runoff results of multiple sub-basins to generate multiple feature tensors; step 2.2, structurally adjusting each feature tensor, and recording the runoff result corresponding to each feature tensor after the structural adjustment as the first runoff result.

[0044] Specifically, in a further embodiment, in order to realize the learning of the flow generation relationship between different sub-basins, based on the characteristics of convolution calculation, the learning of the relationship between different sub-basins is usually local, and the interaction of flood waves between multiple sub-basins is not fully considered, and it is difficult to accurately describe the complex nonlinear confluence process in small basins. Based on the above reasons, the present application adopts feature mapping conversion to adjust the tensor structure input to the convolutional long short-term memory network to realize the learning of the correlation relationship between any two sub-basins. Specifically, in a further embodiment, the features of the flow generation results of multiple sub-basins obtained in step 1 are extracted to generate multiple feature tensors; each feature tensor is structurally adjusted, and the flow generation result corresponding to each feature tensor after the structural adjustment is recorded as the first flow generation result.

[0045] Furthermore, in order to ensure that high-precision hydrological forecast results are obtained under limited data conditions, the present application proposes to perform attention calculation on the runoff results of multiple sub-basins before feature mapping conversion.

[0046] Specifically, in traditional runoff simulation applications, the attention mechanism module adopts a method of calculating attention on time and space separately. However, this method is prone to cause the loss of spatiotemporal correlation information. Therefore, the attention calculation of this application calculates time and space at the same time to ensure that the interaction between the spatial grids in the time series can be taken into account, so as to capture more important spatiotemporal correlation information.

[0047] Furthermore, attention calculation is performed on the runoff results of multiple sub-basins, including:

[0048] Convolution operations are performed on the runoff results of multiple sub-basins respectively, and multiple groups of feature sequences are generated according to the runoff results of multiple sub-basins; nonlinear spatiotemporal attention is assigned to each group of feature sequences to obtain the attention score of each group of feature sequences; weighted calculation is performed on each group of sequence features according to the attention score to generate multiple groups of weighted feature representations.

[0049] Specifically, in this embodiment, Figure 2As shown, a convolution operation is applied to the flow feature sequence of each sub-basin input. Specifically, the convolution kernel size used in this embodiment is 1×1, and the convolution operation will perform weighted summation on the features at each time step without crossing the time step. At the same time, we set the step size of the sliding window to 1 to ensure that the convolution operation can traverse the entire sequence without missing any information of the time step. Then, in order to further enhance the ability to capture key information, we connect the attention layer with the layer after the convolution operation. The attention layer can give the model the ability to dynamically adjust the importance weights of different time steps or features, thereby helping the model to more accurately understand the spatiotemporal relationship in the data. It is worth noting that when connecting the attention layer with the layer after the convolution operation, we pay special attention to maintaining the spatiotemporal structure of the data. In this way, the time order in the original data and the spatial relationship between sub-basins will not be disrupted or destroyed during data transmission. It can ensure that the hydrological forecasting system does not lose key information when extracting features, thereby improving the prediction performance of the hydrological forecasting system.

[0050] Specifically, nonlinear spatiotemporal attention scoring is performed on each set of feature sequences to obtain the attention score of each set of feature sequences;

[0051] Among them, nonlinear spatiotemporal attention is assigned to each set of feature sequences, specifically including:

[0052] Determine all data points in each set of feature sequences according to a preset time and a preset spatial position;

[0053] The spatiotemporal attention score of a data point is determined according to its feature value, and the set of spatiotemporal attention scores of all data points is recorded as the nonlinear spatiotemporal attention score of each set of feature sequences.

[0054] In this embodiment, the nonlinear spatiotemporal attention score calculation formula for each data point is:

[0055]

[0056] In the formula,

[0057] mark(f ij ) is the spatiotemporal attention score calculated for sample j at time i;

[0058] f is the data point in each feature sequence;

[0059] i and j are the temporal and spatial index parameters of the data points respectively.

[0060] A weighted calculation is performed on each group of sequence features according to the attention score to generate multiple groups of weighted feature representations.

[0061] Each group of sequence features is weighted according to the attention score to generate multiple groups of weighted feature representations, specifically including: obtaining the sum of the attention scores of all spatial positions at each time point in each group of sequence features, and recording it as the time point attention score; determining the probability value of the data point according to the spatiotemporal attention score of each data point and the time point attention score of its time point; and each group of sequence features is weighted according to the probability value to generate multiple groups of weighted feature representations.

[0062] Specifically, in this embodiment, the attention score is firstly processed probabilistically, and the calculation formula is as follows:

[0063]

[0064] Where:

[0065] norm(f ij ) is the result after the probabilistic attention score of sample j at time i;

[0066] e is a natural constant;

[0067] n is the number of grids in the spatial sequence;

[0068] Subsequently, the probabilistic results are weighted with the sequence features to enhance the focus on important time steps.

[0069] In a further embodiment, after performing attention calculation on the flow generation results of the plurality of sub-watersheds, extracting feature tensors of a plurality of groups of weighted feature representations, and performing structural adjustment on each feature tensor;

[0070] The structurally adjusted feature tensor is mapped and transformed according to the parity of the time index, and the feature representation corresponding to the mapped feature tensor is recorded as the first flow generation result.

[0071] Specifically, in this embodiment, the calculation process of the feature map is:

[0072]

[0073] λ=(i / 2+k)mods

[0074] In the formula,

[0075] map is the result of feature mapping;

[0076] a is the feature input to the feature mapping layer, that is, the result of attention calculation;

[0077] s is the sample index parameter, which represents different sub-basin samples;

[0078] i is the time index parameter;

[0079] k and λ are the spatial index parameters of the two-dimensional matrix;

[0080] mod is the modulus calculation symbol.

[0081] Step 3: Input the multiple first runoff results into a convolutional long short-term memory neural network model (Convolutional Long Short-Term Memory Neural Network, ConvLSTM) to perform runoff calculation, obtain the runoff results of the preset area, and perform hydrological forecasting based on the runoff results.

[0082] Specifically, in this embodiment, the convolutional long short-term memory neural network model is calculated by the following steps:

[0083] Perform a two-dimensional convolution operation on the hidden state of the previous time step and the current input data. The specific calculation formula is:

[0084]

[0085] In the formula,

[0086] O (i，j) Represents the pixel value of the coordinate position in the output image O;

[0087] x and y are the horizontal and vertical coordinates in the input image respectively;

[0088] ix and jy represent the positions of the convolution kernel on the input image respectively;

[0089] h t-1 Represents the output of the model at the previous moment, i.e., moment t-1;

[0090] x t is the current input data.

[0091] Update the current unit state through the forget gate, input gate and output gate, selectively remember or forget historical information, and calculate it specifically through the following method:

[0092] The model’s “forget gate” forgets information that is likely to lead to incorrect predictions. The forget gate matrix f at time t is t The calculation formula is:

[0093] f t =sigmod(W f ·O+b f )

[0094] In the formula,

[0095] O is the matrix after two-dimensional convolution;

[0096] W f is the parameter matrix;

[0097] b f For bias top.

[0098] The model's "input gate" is calculated using the following formula:

[0099] i t =sigmod(W i ·O+b i )

[0100]

[0101]

[0102] In the formula,

[0103] i t represents the input gate matrix;

[0104] W i Represents the parameter matrix of this step;

[0105] b i is the bias term;

[0106] Indicates the current memory;

[0107] W C A weight matrix representing the input information;

[0108] b C is the bias term;

[0109] C t-1 Represents the information at time t-1;

[0110] C t Indicates the current status.

[0111] The output of each time step is determined by the current cell state and the output gate:

[0112] The new cell state is passed through the tanh network layer and multiplied by the part to be output through the "output gate" to obtain the output result. The calculation formula is as follows:

[0113] o t =sigmod(W o ·[h t-1 ,x t ]+b o )

[0114] h t =o t *tanh(C t )

[0115] In the formula,

[0116] o t represents the output gate matrix;

[0117] W o Represents the parameter matrix of this step;

[0118] b o is the bias term;

[0119] h t Represents the output at time t, that is, the runoff error value at that moment.

[0120] Furthermore, the convolutional long short-term memory neural network model is optimized and trained using the error direction propagation algorithm.

[0121] Furthermore, the loss function of the convolutional long short-term memory neural network model is the mean square error;

[0122] The convolutional long short-term memory neural network model saves the best model through early stopping.

[0123] Specifically, in this embodiment, the learning rate of the convolutional long short-term memory neural network model is set to 0.001, the maximum number of iterations is 2000, and the model saves the optimal weights through the early stopping method. The early stopping cycle is 200 to ensure that the model avoids overfitting during training and enhances the generalization ability of the model.

[0124] Furthermore, this embodiment also includes evaluating the model in the hydrological forecasting method, specifically including evaluating the overall regression degree of the model based on the Nash efficiency coefficient and the correlation coefficient; evaluating the total water volume error during the calculation period based on the relative error; and evaluating the prediction performance of the model based on the mean absolute error and the mean square error.

[0125] Embodiment 2:

[0126] A hydrological forecasting system integrating the spatial and temporal relationship of runoff, including the runoff generation module of the Xin'anjiang model and the convolutional long short-term memory neural network model;

[0127] The runoff module of the Xin'anjiang model is used to calculate the runoff results of multiple sub-basins within a preset area;

[0128] The convolutional long short-term memory neural network model couples the attention mechanism with the feature mapping method, which is used to convert the runoff results of multiple sub-basins into multiple first runoff results, and perform confluence calculation based on the multiple first runoff results.

[0129] Specifically, in this embodiment, we combine the runoff module of the Xin'anjiang model with a convolutional long short-term memory neural network model that integrates an attention mechanism and a specific feature mapping method to form a hydrological forecasting system, which is called the improved Xin'anjiang model.

[0130] This paper proposes a specific feature mapping method by fully considering the spatiotemporal relationship of the confluence process. This method can accurately capture the interaction of flood waves between sub-basins and effectively learn the correlation between any two sub-basins. This technological breakthrough greatly improves the model's ability to understand and simulate flood processes, providing a more accurate tool for flood forecasting.

[0131] The present invention uses attention computing technology to focus on key spatiotemporal features, thereby further optimizing the performance of the improved Xinanjiang model. The application of this technology enables the model to perform well in capturing spatiotemporal dependencies within the basin, and maintain high stability and reliability even when data is limited.

[0132] The present invention combines convolutional long short-term memory neural network to model runoff sequence, further enhancing the model's ability to capture complex spatiotemporal relationships. This combination not only improves simulation accuracy, but also provides more reliable data support for watershed management and flood control decision-making.

[0133] The forecasting method of the present invention has achieved remarkable results in the application of small and medium-sized watersheds. It not only improves the accuracy of hydrological forecasts, but also can be widely used in various watershed environments due to its strong adaptability. This has brought significant social and economic benefits to the fields of water resources management, flood prevention and control, and ecological protection, and provided a strong scientific basis for relevant decision-making.

[0134] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A hydrological forecasting method integrating the temporal and spatial relationship of runoff, characterized in that: include: Step 1: Calculate the runoff results of multiple sub-basins in the preset area according to the runoff module of the Xin'anjiang model; Step 2, converting the runoff generation results of the plurality of sub-watersheds into a plurality of first runoff generation results respectively; Step 3: Input the plurality of the first runoff generation results into a convolutional long short-term memory neural network model to perform runoff calculation, obtain the runoff results of the preset area, and perform hydrological forecasting based on the runoff results.

2. The hydrological forecasting method according to claim 1, characterized in that: Step 2 specifically includes: Step 2.1, extracting the features of the runoff results of the plurality of sub-basins to generate a plurality of feature tensors; Step 2.2: perform structural adjustment on each feature tensor, and record the flow generation result corresponding to each feature tensor after the structural adjustment as the first flow generation result.

3. The hydrological forecasting method according to claim 2, characterized in that: Before the feature map conversion, the method further includes: Attention calculation is performed on the flow generation results of the multiple sub-basins.

4. The hydrological forecasting method according to claim 3, characterized in that: Attention calculation is performed on the flow generation results of the plurality of sub-basins, specifically including: Performing convolution operations on the runoff generation results of the plurality of sub-basins respectively, and generating a plurality of groups of feature sequences according to the runoff generation results of the plurality of sub-basins; Performing nonlinear spatiotemporal attention scoring on each group of feature sequences to obtain an attention score for each group of feature sequences; A weighted calculation is performed on each group of sequence features according to the attention score to generate multiple groups of weighted feature representations.

5. The hydrological forecasting method according to claim 4, characterized in that: Nonlinear spatiotemporal attention scoring is performed on each set of feature sequences, specifically including: Determine all data points in each group of the feature sequences according to a preset time and a preset spatial position; The spatiotemporal attention score of the data point is determined according to its characteristic value, and the set of the spatiotemporal attention scores of all data points is recorded as the nonlinear spatiotemporal attention score of each group of the feature sequences.

6. The hydrological forecasting method according to claim 5, characterized in that: A weighted calculation is performed on each group of sequence features according to the attention score to generate multiple groups of weighted feature representations, including: Obtaining the sum of the attention scores of all spatial positions at each time point in each group of sequence features, and recording it as the time point attention score; Determine the probability value of each data point according to the spatiotemporal attention score of each data point and the time point attention score of the time point at which it is located; A weighted calculation is performed on each group of sequence features according to the probability value to generate multiple groups of weighted feature representations.

7. The hydrological forecasting method according to any one of claims 3 to 6, characterized in that: Step 2 specifically includes: Extracting feature tensors of multiple groups of weighted feature representations, and performing structural adjustment on each of the feature tensors; The structurally adjusted feature tensor is mapped and transformed according to the parity of the time index, and the feature representation corresponding to the feature tensor after the mapping transformation is recorded as the first flow generation result.

8. The hydrological forecasting method according to claim 1, characterized in that: The convolutional long short-term memory neural network model is optimized and trained by an error direction propagation algorithm.

9. The hydrological forecasting method according to claim 1, characterized in that: The loss function of the convolutional long short-term memory neural network model is mean square error; The convolutional long short-term memory neural network model saves the best model through the early stopping method.

10. A hydrological forecasting system integrating the temporal and spatial relationship of runoff, characterized in that: Including the runoff module of the Xinanjiang model and the convolutional long short-term memory neural network model; The runoff generation module of the Xin'anjiang model is used to calculate the runoff generation results of multiple sub-basins in a preset area; The convolutional long short-term memory neural network model couples the attention mechanism with the feature mapping method, and is used to convert the runoff generation results of the multiple sub-basins into multiple first runoff generation results, and perform confluence calculation based on the multiple first runoff generation results.

Images