Multi-source time sequence decomposition and fusion hydraulic structure behavior prediction method

Abstract

The invention discloses a multi-source time sequence decomposition and fusion hydraulic structure behavior prediction method, which comprises the steps of 1, acquiring historical time sequence monitoring data of a hydraulic structure, including behavior parameters and environmental quantity parameters, the behavior parameters including dam body settlement and seepage flow, and the environmental quantity parameters including dam body water level and rainfall; 2, preprocessing the acquired monitoring data, and constructing a hydraulic structure behavior model input data set; 3, determining an optimal decomposition period of an input data set of the hydraulic structure performance model; 4, according to the determined optimal decomposition period, decomposing the preprocessed hydraulic structure performance model input data set into three component parameters according to an original time sequence, wherein the three component parameters are respectively a trend component parameter, a periodic component parameter and a residual component parameter; and 5, superposing the predicted values of the three component parameters according to time points to obtain a final performance predicted value of the hydraulic structure. According to the method, the prediction precision is remarkably improved, and the stability and reliability of long-term prediction are remarkably improved.

Description

Technical Field

[0001] This invention relates to a method for predicting the behavior of hydraulic structures using multi-source time-series decomposition and fusion. Background Technology

[0002] Hydraulic structures, as key infrastructure for national water resource regulation, flood control and disaster reduction, and clean energy power generation, are crucial to the safety of people's lives and property and the overall economic and social development. Deformation monitoring is one of the most intuitive and essential technical means to assess the structural behavior of hydraulic structures. Through continuous observation and analysis of parameters such as displacement and seepage, the inherent evolutionary laws and potential risks of hydraulic structures can be effectively revealed. With the widespread application of sensing technology, the Internet of Things, and automated data acquisition systems, modern hydraulic structure safety monitoring has entered an era of massive, multi-source data. However, this monitoring data usually exhibits strong nonlinearity, non-stationarity, temporal correlation, and high noise levels (affected by the coupling of multiple factors such as ambient temperature, water level changes, material creep, and measurement errors), making it difficult for traditional modeling methods to accurately capture its complex characteristics and achieve high-precision predictions.

[0003] In existing technologies, there are two main approaches to predicting the performance of hydraulic structures: the first is prediction methods based on a single machine learning model. These methods directly utilize historical monitoring data to train models such as Support Vector Regression (SVM), Backpropagation (BP) neural networks, or Long Short-Term Memory (LSTM) networks, aiming to establish a nonlinear mapping from input factors (such as water level, temperature, and time) to deformation. For example, Ou Bin et al. attempted to leverage the advantage of LSTM networks in processing the influence of earlier information in time-series samples to construct a concrete dam deformation prediction model; Luo Lu et al. first used minimum absolute value contraction and selection operators to screen the optimal influencing factors, and then input them into an LSTM network to establish a prediction model. The advantage of this type of method lies in its unified model structure, but its significant limitations are: for raw monitoring data with low signal-to-noise ratios, the model is easily affected by noise interference, making it difficult to converge to the ideal solution; its prediction accuracy is highly dependent on parameter selection; it lacks generalization ability when dealing with complex nonlinear relationships; and it is prone to getting trapped in local minima.

[0004] The second category is prediction methods based on signal decomposition and integration. To improve data quality, this type of method introduces signal processing techniques to preprocess the original sequence. For example, Empirical Mode Decomposition (EMD) is used to decompose deformation monitoring data of hydraulic structures into several Intrinsic Mode Function (IMF) components to uncover data patterns and characteristics; some studies have also used Singular Spectrum Analysis (SSA) to decompose measured data and obtain subsequences such as trend terms and periodic terms. Subsequently, sub-prediction models are constructed based on the characteristics of each component (e.g., wavelet thresholding is used to denoise the high-frequency IMF component before prediction, or the periodic term is predicted as a single time series). Finally, the prediction results of each component are integrated to obtain the final predicted value. This type of method reduces the modeling difficulty to some extent through a "divide and conquer" strategy. However, its inherent defects are that although the decomposed subsequences are stationary, their physical meaning may be ambiguous, and the sub-models are mostly "loosely coupled," failing to fully consider the intrinsic correlation between components and the dynamic response mechanism to environmental driving factors (such as periodic changes in reservoir water level and temperature field fluctuations). In addition, methods such as EMD have mode aliasing problems, which may affect the decomposition effect. Empirical results show that the prediction accuracy of this type of method still has bottlenecks when dealing with highly volatile and noisy data.

[0005] Furthermore, although current research shows a trend of combining digital twin technology, large-scale automated monitoring systems, and risk diagnosis methods based on multi-source data fusion to achieve more intelligent security monitoring and early warning, these systems or methods mostly focus on data collection, visualization, or preliminary analysis. At the core algorithm level of predictive modeling, they still face common challenges such as how to deeply integrate data-driven approaches with physical mechanisms, how to effectively utilize long-term multi-source data for feature enhancement, and how to improve the robustness and interpretability of models in small-sample, high-noise scenarios. Summary of the Invention

[0006] The purpose of this invention is to propose a multi-source time-series decomposition and fusion method for predicting the behavior of hydraulic structures. This method retains the advantages of decomposition and fusion strategies in handling complex sequences, organically integrates multi-source environmental information and physical mechanisms, and possesses strong noise immunity and accurate prediction performance for various components.

[0007] To achieve the above objectives, the solution of the present invention is as follows:

[0008] A method for predicting the performance of hydraulic structures through multi-source temporal decomposition and fusion, the method comprising the following steps:

[0009] Step 1: Obtain historical time-series monitoring data of hydraulic structures, including performance parameters and environmental parameters. Performance parameters include dam settlement and seepage flow, while environmental parameters include dam water level and rainfall.

[0010] Step 2: Preprocess the acquired monitoring data to construct the input dataset for the hydraulic structure behavior model;

[0011] Step 3: Determine the optimal decomposition period of the input dataset for the behavior model of hydraulic structures;

[0012] Step 4: Based on the determined optimal decomposition period, decompose the preprocessed hydraulic structure behavior model input dataset into three component parameters according to the original time series: trend component parameters, periodic component parameters, and residual component parameters.

[0013] Step 5: Prediction result fusion:

[0014] Input the trend component parameters into the trend term prediction model to obtain the trend component prediction value;

[0015] Input the periodic component parameters into the periodic term prediction model to obtain the predicted value of the periodic component;

[0016] Input the residual component parameters into the residual term prediction model to obtain the predicted values ​​of the residual components;

[0017] The predicted values ​​of the trend component, the periodic component, and the residual component are superimposed at time points to obtain the final predicted value of the hydraulic structure behavior.

[0018] The solution further states that the trend term prediction model is a trend term prediction model that determines the optimal hyperparameter combination by training on the input trend component parameters.

[0019] The periodic term prediction model is a periodic term prediction model that determines the optimal regression function by training with input periodic component parameters.

[0020] The residual prediction model is obtained by training the residual component parameters to determine the optimal regression function.

[0021] The scheme further states that the trend component parameters are the time sequence data of dam settlement and seepage flow, which are used to reflect the long-term development trend of hydraulic structure performance and are mainly affected by material rheology and time aging factors.

[0022] The periodic component parameters are periodic sequence data of dam settlement and seepage flow over time, and periodic sequence data of dam water level and rainfall over time, used to reflect the fluctuations in the behavior of hydraulic structures related to the periodic changes in reservoir water level and temperature environmental factors.

[0023] The residual component parameters are the residual component sequence data after STL decomposition of the optimal periodic parameters, used to reflect the trend and residual fluctuations after the period caused by random noise or sudden events.

[0024] The solution further includes: In step 2, the step of preprocessing the acquired monitoring data to construct the input dataset for the hydraulic structure behavior model is as follows:

[0025] The first step is to clean the acquired monitoring data: apply the Laida criterion to identify and remove outliers; for missing data points, use the average of adjacent data points for secondary interpolation to complete the missing data points;

[0026] The second step is to perform time alignment processing on the time series of hydraulic structure performance and environmental quantity series to ensure that the data points correspond one-to-one on the time axis.

[0027] The third step is to normalize or standardize the data processed in the first and second steps to eliminate the influence of dimensions and form the input dataset for the behavior model of hydraulic structures.

[0028] The solution further includes: In step 3, the step of determining the optimal periodic parameters of the input dataset for the hydraulic structure behavior model is as follows:

[0029] 1. Perform Fast Fourier Transform (FFT) on the time series of environmental quantities with certain periods, such as reservoir water level, analyze their spectral characteristics, and identify the top K dominant periods with significant energy.

[0030] 2. Calculate the autocorrelation function (ACF) of the reservoir water level and environmental quantity series, and analyze its autocorrelation under different lag steps;

[0031] Third, by combining the dominant period identified by FFT and the ACF calculation results, the period with the most significant and stable ACF peak is selected to determine the optimal period parameter for STL decomposition.

[0032] The solution further includes: the trend prediction model adopts an SVR neural network with support vector regression; the periodic prediction model and the residual prediction model respectively adopt an LSTM neural network with long short-term memory.

[0033] The solution further includes: the training of the trend prediction model is the training of an SVR neural network, and the process is as follows:

[0034] Step 1: Obtain the initial training dataset and select a kernel function. Choose a suitable kernel function based on the characteristics of the data, including linear kernel, polynomial kernel, or RBF kernel.

[0035] Step 2: Use grid search, random search, or Bayesian optimization methods, combined with cross-validation, to find the optimal combination of parameters;

[0036] Step 3: Calculate the error between the predicted value and the actual value using the mean square error (MSE).

[0037] Step 4: Train the SVR neural network on the training set, solve for the optimal regression function, and save the regression model with the optimal parameters;

[0038] The training of the periodic term prediction model and the residual term prediction model is the training of an LSTM neural network, and the process is as follows:

[0039] Step 1: Obtain the initial training dataset and initialize the network parameters;

[0040] Step 2: Input the training dataset into the network to start network training. Calculate the error of the measurement points on the output layer based on the network's output value and the expected output value.

[0041] Step 3: Update the weights based on the error on the output layer. If the maximum number of iterations or the required accuracy is reached, proceed to step 4; otherwise, return to step 2.

[0042] Step 4: Once the network training is complete, save the trained LSTM neural network and its corresponding parameters for prediction.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] (1) Significantly improved prediction accuracy: By explicitly separating the trend, periodic and random components of deformation data through STL decomposition, and constructing a dedicated model for each component characteristic for prediction, the model's ability to fit complex deformation patterns and its prediction accuracy are effectively improved.

[0045] (2) Clear physical meaning and strong interpretability: The trend component, periodic component and residual component obtained by decomposition have clear physical meanings (corresponding to long-term rheology, environmental periodic driving and random disturbance respectively), making the prediction results more interpretable and facilitating engineers to understand the dominant factors in the behavior development of hydraulic structures.

[0046] (3) Good long-term forecast stability: The separate modeling method effectively reduces the coupling interference between different components. In particular, the trend component that reflects the long-term development trend is modeled independently, which significantly improves the stability and reliability of long-term forecasts.

[0047] The invention will be further explained in detail below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0048] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0049] A multi-source temporal decomposition and fusion method for predicting the behavior of hydraulic structures, such as Figure 1 As shown, the method steps include:

[0050] Step 1: Obtain historical time-series monitoring data of hydraulic structures, including performance parameters and environmental parameters. Performance parameters include dam deformation, settlement and seepage flow, while environmental parameters include dam water level and rainfall.

[0051] Step 2: Perform data preprocessing on the acquired monitoring data to construct the input dataset for the hydraulic structure behavior model;

[0052] Step 3: Determine the optimal decomposition period T for the input dataset of the hydraulic structure behavior model;

[0053] Step 4: Based on the determined optimal decomposition period T, decompose the preprocessed hydraulic structure behavior model input dataset into three component parameters according to the original time series: trend component parameters, period component parameters, and residual component parameters.

[0054] Step 5: Prediction result fusion:

[0055] Inputting the trend component parameters into the trend term prediction model yields the predicted value T of the trend component. t ;

[0056] The periodic component parameters are input into the periodic term prediction model to obtain the predicted value S of the periodic component. t ;

[0057] The residual component parameters are input into the residual term prediction model to obtain the predicted value R of the residual components. t ;

[0058] Simultaneously, the optimal periodic parameter T is used to perform STL decomposition on the environmental quantity sequence to extract its periodic components, which can be used to assist in the prediction of periodic components.

[0059] The predicted value of the trend component T t Predicted value of periodic component S t and residual component prediction value R t By overlaying the data at different time points, the final predicted values ​​of the hydraulic structure's behavior are obtained. .

[0060] Wherein: the trend term prediction model is a trend term prediction model that determines the optimal hyperparameter combination by training with input trend component parameters;

[0061] The periodic term prediction model is a periodic term prediction model that determines the optimal regression function by training with input periodic component parameters.

[0062] The residual prediction model is obtained by training the residual component parameters to determine the optimal regression function.

[0063] The trend component parameters are the time sequence data of dam settlement and seepage flow, which are used to reflect the long-term development trend of hydraulic structure performance and are mainly affected by material rheology and time aging factors.

[0064] The periodic component parameters are periodic sequence data of dam settlement and seepage flow over time, and periodic sequence data of dam water level and rainfall over time, used to reflect the fluctuations in the behavior of hydraulic structures related to the periodic changes in reservoir water level and temperature environmental factors.

[0065] The residual component parameters are the residual component sequence data after STL decomposition of the optimal periodic parameters, used to reflect the trend and residual fluctuations after the period caused by random noise or sudden events.

[0066] In step 2, the step of preprocessing the acquired monitoring data to construct the input dataset for the hydraulic structure behavior model is as follows:

[0067] The first step is to clean the acquired monitoring data: apply the Laida criterion (3σ criterion) to identify and remove outliers; for missing data points, use the average of adjacent data points for secondary interpolation to complete the missing data points;

[0068] The second step is to perform time alignment processing on the time series of hydraulic structure performance and environmental quantity series to ensure that the data points correspond one-to-one on the time axis and achieve data standardization.

[0069] The third step is to normalize or standardize the data processed in the first and second steps to eliminate the influence of dimensions and form the input dataset for the behavior model of hydraulic structures.

[0070] In step 3, the step of determining the optimal periodic parameter T of the input dataset for the behavior model of hydraulic structures is as follows:

[0071] 1. Perform Fast Fourier Transform (FFT) on the time series of environmental quantities with certain periods, such as reservoir water level, analyze their spectral characteristics, and identify the top K dominant periods with significant energy (such as 12 months, 6 months, etc.).

[0072] 2. Calculate the autocorrelation function (ACF) of the reservoir water level and environmental quantity series, and analyze its autocorrelation under different lag steps (corresponding candidate periods);

[0073] Third, combining the dominant period identified by FFT and the ACF calculation results, select the period with the most significant and stable ACF peak, and determine the optimal period parameter T for STL decomposition (for example, determine the optimal period for reservoir water level as 12 months).

[0074] In this embodiment: the trend term prediction model uses an SVR neural network with support vector regression; the periodic term prediction model and the residual term prediction model use an LSTM neural network with long short-term memory, respectively.

[0075] For Support Vector Regression (SVR) neural network models, the input typically includes historical trend component sequences, and time features (such as cumulative time) can be considered. The model focuses on learning the long-term evolution of the behavior of hydraulic structures. Support Vector Regression (SVR) is a machine learning algorithm based on statistical learning theory, which has significant advantages in handling nonlinear regression tasks. The principle is as follows:

[0076] Set training set ,in These are input features. This is the corresponding output. The goal of SVR is to construct a regression function and find a function. The goal is to predict the output y such that the error between the predicted value and the actual output value obtained by SVR is as small as possible, and the margin is maximized within the allowable error range. The mathematical formula is as follows:

[0077] b (1)

[0078] The optimization objective is:

[0079] (2)

[0080] st:

[0081] (3)

[0082] Where w is the model weight vector; b is the bias term; C is the penalty factor, used to balance model complexity and training error; ε is the insensitive loss parameter; and y is the actual value. This is a predicted value; The slack variable is used to handle samples that do not satisfy the ε-insensitive band. To handle nonlinear regression problems, SVR introduces a kernel function to map the input data to a high-dimensional space, enabling the search for a linear regression model in that space. The core hyperparameters of the model include the regularization coefficient C, the kernel function parameter g, and the kernel function parameter γ. The optimization objective is to minimize the root mean square error (RMSE). The insensitive loss parameter ε has a small impact on the model, so a default value of 0.1 is sufficient. Other parameters are found using methods such as grid search, random search, or Bayesian optimization to find the optimal combination of hyperparameters.

[0083] LSTM, based on Recurrent Neural Networks (RNNs), adds three gate units and a unit state structure to each neuron. The gate structures control information transmission at different times during information transfer, enabling long-term information dependency. The forget gate determines how much old information from the previous unit's output is forgotten. The input gate records the current unit's input and the valid information retained by the previous unit. The output gate uses the information from the first two gates and combines it with the unit state to determine the current unit's output. Activation functions are used to improve the network's nonlinear fitting ability; commonly used activation functions include the Sigmoid function and the Tanh function. The specific calculation process of the entire unit and each gate structure is as follows:

[0084] (1) Output of the previous unit's hidden layer and the input value of the current cell The forget gate output is updated as input to the function, as shown in equation (4):

[0085] (4)

[0086] In the formula, Output for the forget gate. and This is the model weight matrix. For bias terms, It is the Sigmoid activation function, i.e. The result can be mapped to (0,1).

[0087] (2) Output of the previous unit's hidden layer and the input of the current unit The input gate output is updated as the input to the function, as shown in equation (5):

[0088] (5)

[0089] In the formula, Let be the output of the input gate at time t. The input gate combines the current input with long-term memory to mine temporal features in the monitoring sequence. Then, the result is calculated using equation (6). .

[0090] (6)

[0091] (3) Output of the previous unit's hidden layer and the input of the current unit The output gate is updated as the input to the function, as shown in equation (7):

[0092] (7)

[0093] In the formula, This is the output of the output gate. Based on the output of the forget gate... and input gate output The network updates the unit status through formula (8).

[0094] (8)

[0095] In the formula, The state of the cell at time t; This is the Hadamard product (element-wise multiplication of vectors). It is obtained from the output gate. and updated cell state Get the current unit output value As in equation (9):

[0096] (9)

[0097] As can be seen from the above calculation process, the hidden layers of the LSTM model gate structure contain a large number of weight matrices and bias terms, and the hidden layers are connected by fully connected layers to realize the function of the LSTM gate structure. This is also the most important difference from the forward propagation of traditional neural networks. When there is new data input in the data sequence, the output is calculated by combining the current input and the internal state of the previous time step, and the internal state is constantly updated during the calculation process.

[0098] Model parameter settings (preferred): The number of hidden layer neurons should be no less than 32, with 2-3 layers. The activation function within the LSTM unit is commonly tanh; sigmoid or linear activation can be considered for the output layer or other layers. The optimizer uses the Adaptive Moment Estimation (Adam) algorithm, with an initial learning rate of 0.0001. The loss function is the mean squared error (MSE). A dropout layer (with a recommended dropout rate of 0.2) is introduced after the LSTM layers to prevent overfitting. The input dimension should consider the time step and the number of features. The output dimension is set according to the prediction step size (e.g., 5 time steps after prediction, then output 5). Divide the model into training and validation sets (e.g., 80% for training, the remaining 20% ​​for validation), and train the model until the loss function converges; the converged loss function is the optimal regression function.

[0099] In this embodiment: the training of the trend prediction model is the training of the SVR neural network, and the process is as follows:

[0100] Step 1: Obtain the initial training dataset and select a kernel function. Choose a suitable kernel function based on the characteristics of the data, including linear kernel, polynomial kernel, or RBF kernel.

[0101] Step 2: Use grid search, random search, or Bayesian optimization methods, combined with cross-validation, to find the optimal combination of parameters. .

[0102] Step 3: Calculate the error between the predicted value and the actual value using the mean square error (MSE).

[0103] Step 4: Train the SVR neural network on the training set, solve for the optimal regression function, and save the regression model with the optimal parameters;

[0104] The training of the periodic term prediction model and the residual term prediction model is the training of an LSTM neural network, and the process is as follows:

[0105] Step 1: Obtain the initial training dataset and initialize the network parameters;

[0106] Step 2: Input the training dataset into the network to start network training. Calculate the error of the measurement points on the output layer based on the network's output value and the expected output value.

[0107] Step 3: Update the weights based on the error on the output layer. If the maximum number of iterations or the required accuracy is reached, proceed to step 4; otherwise, return to step 2.

[0108] Step 4: Once the network training is complete, save the trained LSTM neural network and its corresponding parameters for prediction.

[0109] Compared with existing technologies, the above-described multi-source time-series decomposition and fusion method for predicting the performance of hydraulic structures has the following advantages:

[0110] (1) Significantly improved prediction accuracy: By explicitly separating the trend, periodic and random components of deformation data through STL decomposition, and constructing a dedicated model for each component characteristic for prediction, the model's ability to fit complex deformation patterns and its prediction accuracy are effectively improved.

[0111] (2) Clear physical meaning and strong interpretability: The trend component, periodic component and residual component obtained by decomposition have clear physical meanings (corresponding to long-term rheology, environmental periodic driving and random disturbance respectively), making the prediction results more interpretable and facilitating engineers to understand the dominant factors in the behavior development of hydraulic structures.

[0112] (3) Good long-term forecast stability: The separate modeling method effectively reduces the coupling interference between different components. In particular, the trend component that reflects the long-term development trend is modeled independently, which significantly improves the stability and reliability of long-term forecasts.

Claims

1. A multi-source time series decomposition fusion method for predicting the behavior of hydraulic structures, characterized in that, The method steps include: Step 1, obtaining historical time series monitoring data of hydraulic structures, including performance parameters and environmental quantity parameters, the performance parameters including dam settlement and seepage flow, and the environmental quantity parameters including dam water level and rainfall; Step 2, preprocessing the obtained monitoring data to construct a hydraulic structure performance model input data set; Step 3, determining the optimal decomposition period of the hydraulic structure performance model input data set; Step 4, decomposing the preprocessed hydraulic structure performance model input data set into three component parameters according to the determined optimal decomposition period, namely trend component parameters, periodic component parameters and residual component parameters; Step 5, result fusion: inputting the trend component parameters into a trend item prediction model to obtain trend component prediction values; inputting the periodic component parameters into a periodic item prediction model to obtain periodic component prediction values; inputting the residual component parameters into a residual item prediction model to obtain residual component prediction values; superimposing the trend component prediction values, the periodic component prediction values and the residual component prediction values according to time points to obtain final hydraulic structure performance prediction values.

2. The hydraulic structure performance prediction method according to claim 1, characterized in that: the trend item prediction model is a trend item prediction model with an optimal hyperparameter combination determined by inputting trend component parameters for training; the periodic item prediction model is a periodic item prediction model with an optimal regression function determined by inputting periodic component parameters for training; the residual item prediction model is a residual item prediction model obtained by inputting residual component parameters for training to determine an optimal regression function.

3. The hydraulic structure performance prediction method according to claim 1 or 2, characterized in that: the trend component parameters are sequence data of dam settlement and seepage flow changing with time, used to reflect the long-term development trend of the hydraulic structure performance, affected by material rheology and aging factors; the periodic component parameters are periodic sequence data of dam settlement and seepage flow changing with time and periodic sequence data of dam water level and rainfall changing with time, used to reflect fluctuations in the hydraulic structure performance related to periodic changes in reservoir water level and temperature environmental factors; the residual component parameters are residual component sequence data after STL decomposition of the optimal periodic parameters, used to reflect the remaining fluctuations after the trend and the periodicity caused by random noise or sudden events.

4. The method of predicting the behavior of a hydraulic structure according to claim 1, characterized in that, In step 2, the step of preprocessing the obtained monitoring data to construct a hydraulic structure performance model input data set is: First step, cleaning the obtained monitoring data: applying the Ralda criterion to identify and eliminate outliers; for missing data points, using the average value of adjacent data points for quadratic interpolation completion; Second step, time alignment processing of the hydraulic structure performance time series and environmental quantity series to ensure that the data points correspond one by one on the time axis; Third step, normalizing or standardizing the data processed in the first and second steps to eliminate the influence of dimension, forming a hydraulic structure performance model input data set.

5. The method of predicting the behavior of a hydraulic structure according to claim 1, wherein In step 3, the step of determining the optimal periodic parameter of the hydraulic structure performance model input data set is: Firstly, the FFT is used to analyze the frequency spectrum characteristics of the time series of reservoir water level and other environmental variables with certain periodicity, and the first K dominant periods with significant energy are identified. Secondly, the ACF of the time series of reservoir water level and environmental variables is calculated, and the autocorrelation at different lag steps is analyzed. Thirdly, the optimal period parameter for STL decomposition is determined by selecting the period with the most significant and stable peak in the ACF, combined with the results of FFT identification and ACF calculation.

6. The method of predicting the behavior of a hydraulic structure according to claim 2, wherein The trend prediction model uses the SVR neural network, and the period prediction model and the residual prediction model use the LSTM neural network.

7. The method according to claim 6, wherein the trend prediction model is trained by using the SVR neural network, and the process is as follows: Step 1: Obtain the initial training data set, and select the kernel function according to the characteristics of the data, including linear kernel, polynomial kernel or RBF kernel. Step 2: Use grid search, random search or Bayesian optimization method to find the best parameter combination combined with cross-validation. Step 3: Calculate the error between the predicted value and the actual value by using the MSE. Step 4: Train the SVR neural network on the training set, solve the optimal regression function, and save the regression model with the optimal parameters. The training of the period prediction model and the residual prediction model is the training of the LSTM neural network, and the process is as follows: Step 1: Obtain the initial training data set and initialize the network parameters. Step 2: Input the training data set into the network and start the network training, calculate the error of the measuring point on the output layer according to the output value and the expected output value of the network. Step 3: Update the weight value according to the error on the output layer, and if the maximum iteration number is reached or the accuracy requirement is met, go to step 4, otherwise return to step 2. Step 4: The network training is completed, and the trained LSTM neural network and its corresponding parameters are saved for prediction. ​

Images