A Stochastic Simulation Method and System for Multi-Station Runoff Processes Based on an Improved Variational Autoencoder

Abstract

This invention discloses a method and system for stochastic simulation of multi-station runoff processes based on an improved variational autoencoder. The method includes the following steps: acquiring three-dimensional runoff data from hydrological stations; training a variational autoencoder model using the three-dimensional runoff data, wherein the variational autoencoder model includes an encoder and a decoder, the encoder converting the input three-dimensional runoff data into latent space data, and the decoder converting the latent space data into simulated data; calculating the error between the simulated data and the input three-dimensional runoff data, as well as the divergence between the input three-dimensional runoff data and the latent space data during training, and determining the parameters of the variational autoencoder model corresponding to the minimum error and divergence; using the three-dimensional runoff data from the hydrological stations as input, generating new simulated data using the trained variational autoencoder model, thereby achieving stochastic simulation of the runoff process. This invention solves the problems of poor accuracy and low efficiency in stochastic runoff simulation.

Description

Technical Field

[0001] This invention belongs to the technical field of hydrological stochastic simulation, and more specifically, relates to a method and system for stochastic simulation of multi-station runoff processes based on an improved variational autoencoder. Background Technology

[0002] Hydrological stochastic simulation is a hydrological time series analysis method that uses given hydrological time series to build models and generate a large number of artificial time series. Its aim is to simulate the randomness of hydrological phenomena to aid in the prediction and management of water resources. Runoff process stochastic simulation is a type of hydrological stochastic simulation. It uses historical runoff processes to build models, simulate the randomness of runoff processes, and generate a large number of simulated runoff processes. This provides support for scientific watershed water resources planning and management, and joint scheduling of reservoir groups. It can adapt to the precise scheduling of different types of runoff processes in the future, meeting the requirements of the new stage of smart water conservancy construction for scientific and precise water conservancy decision-making and management. Single-station runoff process stochastic simulation methods such as seasonal autoregression are relatively mature, but these methods cannot well maintain the spatial distribution characteristics of multi-station runoff processes. Existing multi-station runoff process stochastic simulation methods based on Coupla functions and Gaussian regression can maintain the spatial distribution characteristics of multi-station runoff processes, but their simulation efficiency is low, simulation errors are large, and they require pre-assuming runoff edge distributions, which limits further improvement in the simulation accuracy of these methods.

[0003] In summary, existing stochastic simulation methods for runoff processes have at least the following shortcomings or require improvement:

[0004] (1) The stochastic simulation method for single-station runoff processes cannot well maintain the spatial distribution characteristics of multi-station runoff processes. Although the existing stochastic simulation method for multi-station runoff processes can maintain the spatial distribution characteristics of multi-station runoff processes, the efficiency of stochastic simulation decreases significantly as the number of simulated hydrological stations and the simulation period increase.

[0005] (2) Existing stochastic simulation methods for multi-station runoff processes require prior assumptions about the marginal distribution of the runoff process. Although they can maintain the spatial correlation of the runoff process to a certain extent, the spatial correlation between the simulated multi-station runoff process and the historical runoff process is quite different, and to a certain extent, they limit the simulation accuracy of the runoff process.

[0006] Therefore, there is an urgent need for a method to simulate runoff processes that can solve the above problems. Summary of the Invention

[0007] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a multi-station stochastic simulation method and system for runoff processes based on an improved variational autoencoder, solving the problems of poor accuracy and low efficiency in stochastic simulation of runoff processes.

[0008] To achieve the above objectives, according to one aspect of the present invention, a stochastic simulation method for multi-station runoff processes based on an improved variational autoencoder is provided, the method comprising the following steps:

[0009] Obtain three-dimensional runoff data from hydrological stations. This three-dimensional runoff data includes historical runoff values ​​corresponding to three dimensions: the total number of hydrological stations, the total number of stations per year, and the total number of time periods included in a year.

[0010] A variational autoencoder model is trained using the three-dimensional runoff data. The input of the variational autoencoder model is the three-dimensional runoff data, and the output is the generated simulated data. The variational autoencoder model includes an encoder and a decoder. The encoder is used to convert the input three-dimensional runoff data into latent space data, and the decoder is used to convert the latent space data into simulated data. The encoder includes an input layer, a convolutional layer, an activation layer, a fully connected layer, an attention layer, and a sampling layer connected in sequence. The convolutional layer and the activation layer are used to extract feature information from the input data of the input layer, and the attention layer is used to assign weights to the input from the fully connected layer.

[0011] During training, the error between the simulated data and the input 3D runoff data, as well as the divergence between the input 3D runoff data and the latent space data, are calculated. The parameters of the variational autoencoder model corresponding to the minimum error and divergence are the final determined parameters.

[0012] Using three-dimensional runoff data from hydrological stations as input, new simulation data is generated by the trained variational autoencoder model. This new simulation data is the required simulation data, thereby realizing the stochastic simulation of the runoff process.

[0013] More preferably, the decoder includes an input layer, a reconstruction layer, a transposed convolutional layer, an activation layer, and an output layer connected in sequence. The reconstruction layer is used to project and reconstruct the data input from the input layer, and the transposed convolutional layer and the activation layer are used to decode the data output from the reconstruction layer into simulated data with the same dimension as the data input to the variational autoencoder model.

[0014] More preferably, the encoder has two convolutional layers and two activation layers, which are connected alternately.

[0015] More preferably, the decoder has three transposed convolutional layers and three activation layers, with the transposed convolutional layers and activation layers connected alternately; the two activation layers in the encoder both use the ReLU function as the activation function, and the first two activation layers in the decoder use the ReLU function as the activation function, while the third activation layer uses the Sigmoid function as the activation function.

[0016] More preferably, the calculation process of the attention layer includes the following steps:

[0017] The intermediate matrices R, U, and V are computed using the input from the fully connected layer;

[0018] Calculate the attention score matrix using R and U;

[0019] The output matrix of the attention layer is calculated using the attention score matrix and the intermediate matrix V.

[0020] More preferably, the intermediate matrix is ​​calculated according to the following formula:

[0021] R = W R I

[0022] U = W U I

[0023] V = W V I

[0024] The attention score matrix is ​​calculated according to the following formula:

[0025] A = softmax(U T R)

[0026] The attention layer output matrix is ​​calculated according to the following formula:

[0027] O = VA

[0028] Among them, W R W U and W V All of these are parameters of the attention layer, where I is the input to the attention layer, A is the attention score matrix, and O is the output matrix of the attention layer.

[0029] More preferably, before inputting the three-dimensional runoff data of the hydrological station into the variational autoencoder model, the three-dimensional runoff data of the hydrological station is normalized using the max-min normalization method.

[0030] More preferably, after training the variational autoencoder model, the trained variational autoencoder model is verified and adjusted according to the following steps:

[0031] The three-dimensional runoff data is input into the variational autoencoder model to obtain multiple sets of simulation data;

[0032] The simulation data is then denormalized.

[0033] Calculate the statistical values ​​of the simulated data after inverse normalization, and adjust the parameters in the variational autoencoder model according to the calculated statistical values ​​until the obtained statistical values ​​meet the preset threshold requirements. The statistical values ​​include the mean, coefficient of variation, maximum value and Pearson correlation coefficient.

[0034] More preferably, the error is a mean square error, and the divergence is a relative entropy.

[0035] According to another aspect of the present invention, a stochastic simulation system for multi-station runoff processes based on an improved variational autoencoder is provided. The system includes an actuator for performing the aforementioned stochastic simulation method for multi-station runoff processes based on an improved variational autoencoder.

[0036] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art:

[0037] 1. To address the issue of low simulation accuracy in existing stochastic simulation methods for multi-station runoff processes, this invention constructs a convolutional variational autoencoder model coupled with a self-attention mechanism. This model does not require prior assumptions about the marginal distribution of the runoff process. It achieves multi-station runoff process simulation by learning the spatiotemporal distribution characteristics of the runoff process, which not only improves the simulation accuracy of the runoff process but also improves the simulation accuracy of the spatial correlation of the runoff process to a certain extent.

[0038] 2. To address the problem of low simulation efficiency in existing stochastic simulation methods for multi-station runoff processes, this invention proposes a stochastic simulation method for multi-station runoff processes based on an improved variational autoencoder by introducing deep learning technology. The trained variational autoencoder model can quickly simulate multi-station runoff processes, and the model's simulation efficiency does not decrease significantly as the number of simulated hydrological stations and the simulation period increase.

[0039] 3. This invention verifies the model after training the convolutional variational autoencoder model with coupled self-attention mechanism, and then adjusts the model parameters again through the verification process, thereby realizing two adjustments to the parameters of the convolutional variational autoencoder model with coupled self-attention mechanism, and improving the simulation accuracy of the model. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating a stochastic simulation method for multi-station runoff processes based on an improved variational autoencoder, constructed according to a preferred embodiment of the present invention.

[0041] Figure 2 This is a structural diagram of a convolutional variational autoencoder model with a coupled self-attention mechanism constructed according to a preferred embodiment of the present invention.

[0042] Figure 3The results are the verification results of the statistical parameters of the Xiangjiaba, Gaochang, Tingzikou, and Three Gorges runoff processes simulated by the convolutional variational autoencoder model with coupled self-attention mechanism constructed according to the preferred embodiment of the present invention. Among them, (a) is the simulated average daily flow of the Xiangjiaba runoff process, (b) is the simulated coefficient of variation of the daily flow of the Xiangjiaba runoff process, (c) is the simulated maximum daily flow of the Xiangjiaba runoff process, (d) is the simulated average daily flow of the Gaochang runoff process, (e) is the simulated coefficient of variation of the daily flow of the Gaochang runoff process, (f) is the simulated maximum daily flow of the Gaochang runoff process, (g) is the simulated average daily flow of the Tingzikou runoff process, (h) is the simulated coefficient of variation of the daily flow of the Tingzikou runoff process, (i) is the simulated maximum daily flow of the Tingzikou runoff process, (j) is the simulated average daily flow of the Three Gorges runoff process, (k) is the simulated coefficient of variation of the daily flow of the Three Gorges runoff process, and (l) is the simulated maximum daily flow of the Three Gorges runoff process.

[0043] Figure 4 The results show the verification of the spatial correlation between the Xiangjiaba, Gaochang, Tingzikou and Three Gorges runoff processes simulated by the convolutional variational autoencoder model with coupled self-attention mechanism constructed according to the preferred embodiment of the present invention (the method of the present invention) and the comparative method. Among them, (a) is the Pearson correlation coefficient of daily flow between Xiangjiaba and Three Gorges runoff processes simulated by the method of the present invention, (b) is the Pearson correlation coefficient of daily flow between Xiangjiaba and Three Gorges runoff processes simulated by the comparative method, (c) is the Pearson correlation coefficient of daily flow between Gaochang and Three Gorges runoff processes simulated by the method of the present invention, (d) is the Pearson correlation coefficient of daily flow between Gaochang and Three Gorges runoff processes simulated by the comparative method, (e) is the Pearson correlation coefficient of daily flow between Tingzikou and Three Gorges runoff processes simulated by the method of the present invention, and (f) is the Pearson correlation coefficient of daily flow between Tingzikou and Three Gorges runoff processes simulated by the comparative method.

[0044] Figure 5 The data are runoff events from June 1st to September 30th of the previous 10 years at four hydrological stations: Xiangjiaba, Gaochang, Tingzikou, and Three Gorges, randomly generated by a convolutional variational autoencoder model with a coupled self-attention mechanism constructed according to a preferred embodiment of the present invention. Among them, (a) is the daily flow rate of Xiangjiaba from June 1st to September 30th of the previous 10 years, (b) is the daily flow rate of Gaochang from June 1st to September 30th of the previous 10 years, (c) is the daily flow rate of Tingzikou from June 1st to September 30th of the previous 10 years, and (d) is the daily flow rate of Three Gorges from June 1st to September 30th of the previous 10 years. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0046] This invention provides a stochastic simulation method for multi-station runoff processes based on an improved variational autoencoder. The method constructs a variational autoencoder model, namely a convolutional variational autoencoder model coupled with a self-attention mechanism, to learn the spatiotemporal distribution patterns of historical runoff processes at multiple stations, and randomly generates simulated runoff processes that are similar to the spatiotemporal distribution patterns of historical runoff processes, thereby completing the stochastic simulation of multi-station runoff processes and providing a large number of different types of runoff processes for joint scheduling of reservoir groups.

[0047] Please see Figure 1 The multi-station runoff process stochastic simulation method based on the improved variational autoencoder mainly includes the following steps:

[0048] Step 1 involves constructing the three-dimensional runoff data for the model input based on historical runoff events from multiple stations. This includes the following sub-steps:

[0049] (1.1) The historical runoff processes of multiple stations are organized into runoff values ​​under three dimensions: the total number of hydrological stations, the total number of years, and the total number of time periods included in a year.

[0050] (1.2) The maximum-minimum normalization method is used to normalize the three-dimensional runoff data and construct the three-dimensional runoff data input to the model, denoted as D0.

[0051] Step 2: Construct a convolutional variational autoencoder model with a coupled self-attention mechanism, the structure of which is shown in the figure below. Figure 2 As shown, it consists of two parts: an encoder and a decoder. The encoder takes 3D runoff data D0 as input and generates latent space data D1 with a probability distribution closely resembling that of D0. The decoder takes D1 as input and generates simulated data D2 similar to D1. Constructing a convolutional variational autoencoder with a coupled self-attention mechanism includes the following sub-steps:

[0052] (2.1) Encoder Construction. The encoder consists of an input layer, convolutional layers, activation layers, fully connected layers, attention layers, and sampling layers. The data flow is as follows: training data D0 → input layer → first convolutional layer → first activation layer (using ReLU as the activation function) → second convolutional layer → second activation layer (using ReLU as the activation function) → fully connected layer → attention layer → sampling layer → latent space data D1. The two convolutional layers and the activation layer are used to extract feature information from the 3D runoff data D0. The fully connected layer is used to integrate the extracted features. The attention layer improves the ability to extract data feature information by redistributing the weights between inputs. The sampling layer generates latent space data D1 through sampling operations.

[0053] (2.2) Decoder Construction. The decoder consists of a reconstruction layer, a transposed convolutional layer, an activation layer, and an output layer. The data flow is as follows: latent space data D1 → reconstruction layer → 1st transposed convolutional layer → 1st activation layer (using ReLU as the activation function) → 2nd transposed convolutional layer → 2nd activation layer (using ReLU as the activation function) → 3rd transposed convolutional layer → 3rd activation layer (using Sigmoid as the activation function) → simulated data D2. The reconstruction layer is used to perform data projection and reconstruction operations on the latent space data D1. The three transposed convolutional layers and the activation layer are used to process the output of the reconstruction layer into simulated data D2 with the same data dimension as the three-dimensional runoff data D0.

[0054] Specifically, the attention layer in the encoder employs a self-attention mechanism, and its computation process includes the following sub-steps:

[0055] (2.1.1) Generate intermediate matrices R, U, and V, whose calculation formulas are as follows:

[0056] R = W R I

[0057] U = W U I

[0058] V = W V I

[0059] In the formula: W R W U and W V These are all parameters of the attention layer, obtained through model training; I is the input of the attention layer.

[0060] (2.1.2) Using the obtained intermediate matrices R and U, calculate the attention score matrix A between the inputs. The calculation formula is as follows:

[0061] A = softmax(U T R)

[0062] In the formula: A is the attention score matrix between inputs; softmax is the activation function of the attention layer.

[0063] (2.1.3) Using the obtained attention score matrix A and intermediate matrix V, calculate the output matrix O of the attention layer. The calculation formula is as follows:

[0064] O = VA

[0065] In the formula: O is the output matrix of the attention layer.

[0066] Step 3: Train the convolutional variational autoencoder model with coupled self-attention mechanism using the three-dimensional runoff data D0, including encoder training and decoder training:

[0067] (3.1) Encoder Training: The encoder learns the distribution pattern of the input data using the three-dimensional runoff data D0 as input data, and updates the encoder parameters to make the relative entropy D between the distribution Q of the latent space data D1 and the distribution P of the three-dimensional runoff data D0 equal. KL By minimizing (P||Q), the latent space data D1, which closely resembles the distribution pattern of the three-dimensional runoff data D0, can be deduced.

[0068]

[0069] In the formula: P(x) is the distribution function of the three-dimensional runoff data D0 distribution P; Q(x) is the distribution function of the latent space data D1 distribution Q; X is the data sample space.

[0070] (3.2) Decoder training: The decoder takes the latent space data D1 inferred by the encoder as input data, reconstructs the latent space data D1, and updates the parameters of the decoder to minimize the mean square error (MSE) between the generated simulated data D2 and the three-dimensional runoff data D0, and finally generates simulated data D2 that is similar to the training set data.

[0071]

[0072] In the formula: q m,n,k This refers to the runoff data at time k in year n for the m-th hydrological station in the three-dimensional runoff data D0; s m,n,k The decoder generates simulation data D2 for the runoff data corresponding to the m-th hydrological station at time k in year n; M is the number of hydrological stations; N is the number of years in the historical runoff process; K is the number of time points in the historical runoff process within the year.

[0073] (3.3) Repeat sub-steps (3.1) and (3.2) to continuously train the encoder and decoder until the relative entropy D of distribution Q and P is equal. KLThe mean square error (MSE) of the simulated data D2 and the three-dimensional runoff data D0 reaches the minimum or meets the preset requirements.

[0074] Step four involves validating the trained convolutional variational autoencoder model with coupled self-attention mechanism, including the following sub-steps:

[0075] (4.1) Using the three-dimensional runoff data of the hydrological station as input, the J sets of simulated data are generated randomly by using the trained convolutional variational autoencoder model with coupled self-attention mechanism.

[0076] (4.2) Perform inverse normalization on the J-group simulation data to generate the J-group multi-station simulated runoff process.

[0077] (4.3) The multi-station simulated runoff process simulated by the J group model was tested using statistical parameters and other test indicators.

[0078] Step 5: Using the three-dimensional runoff data from the hydrological station as input, the multi-station runoff process is stochastically simulated using the validated convolutional variational autoencoder model with coupled self-attention mechanism, and then inverse normalization is performed to generate a new multi-station simulated runoff process.

[0079] The present invention also provides a stochastic simulation system for multi-station runoff processes based on an improved variational autoencoder. The system includes an actuator for executing a stochastic simulation method for multi-station runoff processes based on an improved variational autoencoder as described above.

[0080] The present invention will be further described in detail below with reference to examples.

[0081] The procedure for the multi-station runoff process stochastic simulation method based on the improved variational autoencoder provided by this invention is as follows: Figure 1 As shown.

[0082] (1) The historical runoff data from 1960 to 1990 from June 1 to September 30 of four hydrological stations in the upper reaches of the Yangtze River, namely Xiangjiaba, Gaochang, Tingzikou and Three Gorges, were used to construct the model training set, namely the three-dimensional runoff data D1 input to the model.

[0083] (2) Construct a convolutional variational autoencoder model coupled with a self-attention mechanism. The model structure diagram is shown below. Figure 2The model includes an encoder and a decoder. The encoder takes 3D runoff data D0 as input and generates latent space data D1 with a probability distribution close to that of the 3D runoff data D0. The encoder consists of an input layer, a convolutional layer, an activation layer, a fully connected layer, an attention layer, and a sampling layer connected in sequence. The decoder takes latent space data D1 as input and generates simulated data D2 that is similar to the latent space data. The decoder consists of an input layer, a reconstruction layer, a transposed convolutional layer, an activation layer, and an output layer connected in sequence.

[0084] (3) The convolutional variational autoencoder model with coupled self-attention mechanism is trained using three-dimensional runoff data D0, including encoder training and decoder training. Encoder training updates the encoder parameters to make the relative entropy D between the distribution Q of the latent space data D1 and the distribution P of the three-dimensional runoff data D0 equal. KL (P||Q) is minimized; the decoder is trained to minimize the mean square error (MSE) between the generated simulated data D2 and the 3D runoff data D0 by updating the decoder parameters; the encoder and decoder are continuously trained until the relative entropy D of distribution Q and distribution P is minimized. KL The model parameters are determined when the mean square error (MSE) of the simulated data D2 and the three-dimensional runoff data D0 reaches its minimum or meets the preset requirements.

[0085] (4) Using the trained convolutional variational autoencoder model with coupled self-attention mechanism, 30 sets of runoff processes from June 1 to September 30 of 31 consecutive years were randomly generated at four hydrological stations including Xiangjiaba, Gaochang, Tingzikou and Three Gorges.

[0086] The method of this invention was used to randomly generate 30 sets of runoff processes from four hydrological stations over a continuous period of 31 years, from June 1st to September 30th, taking 8 minutes and 3.634 seconds. In contrast, the comparative method, "Multi-station Runoff Simulation Method Based on Coupla Function" published in the journal *Hydrology* by Kang Ling, Guo Jinlei, Zhou Liwei, et al., took 12 minutes and 5.845 seconds. Compared to the comparative method, the method of this invention has higher simulation efficiency, improving it by approximately 33.37%, and this improvement in efficiency gradually increases with the number of hydrological stations to be simulated.

[0087] Then, by generating 30 sets of runoff processes from June 1st to September 30th over 31 consecutive years at four hydrological stations including Xiangjiaba, Gaochang, Tingzikou, and Three Gorges, the simulation accuracy of the convolutional variational autoencoder model coupled with the self-attention mechanism was tested.

[0088] First, statistical parameters such as mean flow, coefficient of variation, and peak flow of the simulated runoff process at multiple stations were tested. The test results of statistical parameters for the simulated runoff process at four hydrological stations—Xiangjiaba, Gaochang, Tingzikou, and Three Gorges—are as follows: Figure 3 As shown, from Figure 3As can be seen from (a) in the figure, the average daily flow rate of the simulated runoff process at Xiangjiaba is generally in good agreement with the measured values, and the trend of change is consistent. Figure 3 From (b) and (c) in the table, it can be seen that the coefficient of variation and maximum value of daily flow in the simulated runoff process at Xiangjiaba are generally close to the measured values, and the trends are basically consistent. Similarly, from Figure 3 From (d) to (l), it can be seen that the average daily flow, coefficient of variation, and maximum value of the simulated runoff process in Gaochang, Tingzikou, and Three Gorges are basically consistent with the measured values.

[0089] The relative errors of the statistical parameters such as the mean, coefficient of variation, and maximum value of the simulated runoff process obtained by the method of this invention are shown in Table 1. Compared with the comparative method, the relative errors of the statistical parameters such as the mean, coefficient of variation, and maximum value of the simulated runoff process obtained by the method of this invention are all lower and can all be kept within 5%. Therefore, the trend of the variation of statistical parameters such as the mean flow, coefficient of variation, and peak flow of the simulated runoff process is consistent with that of the historical runoff process, indicating that the method of this invention can simulate the statistical characteristics of the historical runoff process and has higher simulation accuracy.

[0090] Table 1. Relative errors (%) of statistical parameters such as mean, coefficient of variation, and maximum value of simulated runoff process.

[0091]

[0092] Then, the Pearson correlation between the simulated runoff processes at Xiangjiaba, Gaochang, and Tingzikou stations and the Three Gorges runoff process was tested. The correlation test results between Xiangjiaba, Gaochang, Tingzikou, and the Three Gorges runoff process are as follows: Figure 4 As shown, from Figure 4 As can be seen from (a) and (b) in the figures, compared with the comparative method, the Pearson correlation between the simulated runoff processes of Xiangjiaba and the Three Gorges Dam simulated by the method of this invention is closer to the measured values, and the trends are more consistent. Similarly, from Figure 4 As can be seen from (c) to (f), compared with the comparative method, the Pearson correlations of the simulated runoff processes at Gaochang, Tingzikou, and the Three Gorges Dam using the method of this invention are in better agreement with the measured values. Therefore, the comparative method shows a significant deviation between the simulated Pearson correlation coefficients of different stations and the historical Pearson correlation coefficients, while the simulated Pearson correlation coefficients of different stations using the method of this invention generally follow the same trend as the historical Pearson correlation coefficients. Therefore, the method of this invention can effectively capture the spatiotemporal correlation of inter-station runoff processes, resulting in higher simulation accuracy.

[0093] (5) Using the verified convolutional variational autoencoder model with coupled self-attention mechanism, random simulations of runoff processes at four hydrological stations—Xiangjiaba, Gaochang, Tingzikou, and Three Gorges—are performed. A simulated runoff process of a specified length is randomly generated. In this embodiment, the runoff processes from June 1st to September 30th of 3000 consecutive years at the four hydrological stations are randomly simulated. The runoff processes for the first 10 years are as follows: Figure 5 As shown, from Figure 5 As shown in (a), the flow rate at Xiangjiaba from June 1st to September 30th over the previous 10 years, generated by random simulation, varies over time, and the flow rate variation process differs from year to year. Similarly, from... Figure 5 As shown in (b) to (d), the flow rates of the four hydrological stations (Xiangjiaba, Gaochang, Tingzikou, and Three Gorges) generated by random simulation from June 1st to September 30th over the previous 10 years also changed over time, and the flow rate change process was different in each year. This indicates that the runoff processes of the four hydrological stations (Xiangjiaba, Gaochang, Tingzikou, and Three Gorges) generated by random simulation over the period of 3000 years from June 1st to September 30th have different shapes, which can provide rich runoff samples for joint scheduling of reservoir groups and improve the reliability of joint scheduling schemes of reservoir groups.

[0094] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A stochastic simulation method for multi-station runoff processes based on an improved variational autoencoder, characterized in that, The method includes the following steps: Obtain three-dimensional runoff data from hydrological stations. This three-dimensional runoff data includes historical runoff values ​​corresponding to three dimensions: the total number of hydrological stations, the total number of stations per year, and the total number of time periods included in a year. A variational autoencoder model is trained using the three-dimensional runoff data. The input of the variational autoencoder model is the three-dimensional runoff data, and the output is the generated simulated data. The variational autoencoder model includes an encoder and a decoder. The encoder is used to convert the input three-dimensional runoff data into latent space data, and the decoder is used to convert the latent space data into simulated data. The encoder includes an input layer, a convolutional layer, an activation layer, a fully connected layer, an attention layer, and a sampling layer connected in sequence. The convolutional layer and the activation layer are used to extract feature information from the input data of the input layer, and the attention layer is used to assign weights to the input from the fully connected layer. During training, the error between the simulated data and the input 3D runoff data, as well as the divergence between the input 3D runoff data and the latent space data, are calculated. The parameters of the variational autoencoder model corresponding to the minimum error and divergence are the final determined parameters. Using the three-dimensional runoff data from the hydrological station as input, new simulation data is generated by the trained variational autoencoder model. This new simulation data is the required simulation data, thereby realizing the stochastic simulation of the runoff process.

2. The multi-station runoff process stochastic simulation method based on an improved variational autoencoder as described in claim 1, characterized in that, The decoder includes an input layer, a reconstruction layer, a transposed convolutional layer, an activation layer, and an output layer connected in sequence. The reconstruction layer is used to project and reconstruct the data input from the input layer. The transposed convolutional layer and the activation layer are used to decode the data output from the reconstruction layer into simulated data with the same dimension as the data input to the variational autoencoder model.

3. A stochastic simulation method for multi-station runoff processes based on an improved variational autoencoder as described in claim 1 or 2, characterized in that, The encoder has two convolutional layers and two activation layers, which are connected alternately.

4. The multi-station stochastic simulation method for runoff processes based on an improved variational autoencoder as described in claim 3, characterized in that, The decoder has three transposed convolutional layers and three activation layers, with the transposed convolutional layers and activation layers connected alternately. The encoder uses the ReLU function as the activation function in both activation layers. In the decoder, the first two activation layers use the ReLU function as the activation function, and the third activation layer uses the Sigmoid function as the activation function.

5. A multi-station stochastic simulation method for runoff processes based on an improved variational autoencoder as described in claim 1 or 2, characterized in that, The calculation process for the attention layer includes the following steps: The intermediate matrices R, U, and V are computed using the input from the fully connected layer; Calculate the attention score matrix using R and U; The output matrix of the attention layer is calculated using the attention score matrix and the intermediate matrix V.

6. The multi-station stochastic simulation method for runoff processes based on an improved variational autoencoder as described in claim 5, characterized in that, The intermediate matrix is ​​calculated according to the following formula: R=W R I U=W U I V=W V I The attention score matrix is ​​calculated according to the following formula: A=softmax(U T R) The attention layer output matrix is ​​calculated according to the following formula: O = VA Among them, W R W U and W V All of these are parameters of the attention layer, where I is the input to the attention layer, A is the attention score matrix, and O is the output matrix of the attention layer.

7. The multi-station stochastic simulation method for runoff processes based on an improved variational autoencoder as described in claim 1, characterized in that, Before inputting the three-dimensional runoff data of the hydrological station into the variational autoencoder model, the three-dimensional runoff data of the hydrological station is normalized using the max-min normalization method.

8. The multi-station runoff process stochastic simulation method based on an improved variational autoencoder as described in claim 7, characterized in that, After training the variational autoencoder model, the trained variational autoencoder model is verified and adjusted according to the following steps: Using the three-dimensional runoff data from the hydrological station as input, the three-dimensional runoff data is input into the variational autoencoder model to obtain multiple sets of simulation data; The simulation data is then denormalized. Calculate the statistical values ​​of the simulated data after inverse normalization, and adjust the parameters in the variational autoencoder model according to the calculated statistical values ​​until the obtained statistical values ​​meet the preset threshold requirements. The statistical values ​​include the mean, coefficient of variation, maximum value and Pearson correlation coefficient.

9. A stochastic simulation method for multi-station runoff processes based on an improved variational autoencoder as described in claim 1 or 2, characterized in that, The error is the mean square error, and the divergence is the relative entropy.

10. A stochastic simulation system for multi-station runoff processes based on an improved variational autoencoder, characterized in that, The system includes an actuator for performing a stochastic simulation method for multi-station runoff processes based on an improved variational autoencoder, as described in any one of claims 1-9.

Images