Novel method for improving dam deformation prediction efficiency and precision

Abstract

The invention discloses a novel method for improving dam deformation prediction efficiency and precision, and belongs to the technical field of dam deformation prediction.The method comprises the steps that obtained deformation monitoring data is decomposed into modal components through a variational mode decomposition algorithm, and a modal sample data set is established in combination with environmental quantity monitoring data; based on an Adaboost model, the influence of each piece of environmental quantity monitoring data on the dam deformation effect is considered during weak regression device node segmentation of the Adaboost model, a dynamic characteristic variable weight strategy is introduced, a Huber function is used as a loss function of a weak regression device, targeted improvement is carried out in combination with the characteristics of a dam deformation prediction problem, and a DpAdaBoost model is established; according to the method, a DpAdaBoost model is used as a model, a modal sample data set is used as input, and a GridSampler optimization algorithm is introduced to perform hyper-parameter automatic optimization of the DpAdaBoost model, so that each modal component prediction model of the optimal hyper-parameter is obtained to predict dam deformation in a future time period, an effective solution is provided for improving dam deformation prediction efficiency and precision, and safe operation of the dam is guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of dam deformation prediction, and specifically relates to a new method for improving the efficiency and accuracy of dam deformation prediction. Background Art

[0002] As a key facility for protecting and utilizing water resources, a dam undertakes various functions such as flood control, irrigation, water supply, and power generation, and its safety is of crucial importance. The deformation of a dam is a key factor affecting its operation safety, which may lead to structural damage, dam collapse, decline in flood discharge capacity, and even failure of flood control functions. These situations may threaten the normal operation of the dam and pose a serious threat to the safety of downstream residents and property. Therefore, accurately predicting dam deformation is crucial for maintaining the safe operation of the dam, enabling us to detect and solve potential deformation problems in a timely manner, and ensuring the stability and safety of the dam. At the same time, accurate deformation prediction provides a scientific basis for the maintenance, design, and construction of the dam, helps to develop more effective design schemes, select more suitable building materials, and adopt more advanced construction technologies, thereby reducing the risk of deformation from the source.

[0003] With the rapid development of computers and new-generation information technologies, there are various methods for dam seepage prediction. The most common one is to use machine learning technologies to predict deformation based on various types of massive historical monitoring data of the dam. Currently, many scholars have used various machine learning algorithms to achieve dam deformation prediction. However, due to the historical monitoring data of dam deformation, on the one hand, there will be abnormal error points; on the other hand, dam deformation is comprehensively affected by various environmental factors and internal factors of the dam itself, and the deformation prediction problem is particularly complex. As a result, directly using traditional general models lacks pertinence to the dam deformation prediction problem, and it is particularly difficult to select hyperparameters, resulting in unsatisfactory deformation prediction effects. Summary of the Invention

[0004] The object of the present invention is to address the problem of unsatisfactory accuracy and efficiency of current dam deformation prediction. A new method for improving the efficiency and accuracy of dam deformation prediction is proposed. First, the variational mode decomposition (VMD) algorithm is used to decompose the deformation monitoring data to reduce the complexity of the dam deformation monitoring data. Subsequently, based on the AdaBoost model and in combination with the characteristics of the dam deformation prediction problem, the AdaBoost model is specifically improved to establish the DpAdaBoost model. Finally, the GridSampler optimization algorithm is used to improve the optimization effect of the optimal parameters of the DpAdaBoost model, providing an effective solution for dam deformation prediction. <{

[0005] Specifically, the present invention provides a new method for improving the efficiency and accuracy of dam deformation prediction.

[0006] S1. Obtain the deformation monitoring data and environmental quantity monitoring data of the dam monitoring points, preprocess them, and establish an initial sample data set;

[0007] S2. Based on the initial sample data set, use the variational mode decomposition algorithm to decompose it into several modal components, and combine with the environmental quantity monitoring data to establish a modal sample data set;

[0008] S3. Based on the Adaboost model, introduce a dynamic feature variable weight strategy during the weak regressor node splitting of the Adaboost model and use the Huber function as the loss function of the weak regressor to establish a DpAdaBoost model;

[0009] S4. Use the modal sample data set as the input, each modal component as the output, and introduce the GridSampler optimization algorithm to automatically optimize the hyperparameters of the DpAdaBoost model, obtain the optimal hyperparameters, and thus obtain the prediction models of each modal component with the optimal hyperparameters;

[0010] S5. Input the environmental quantity monitoring data corresponding to the future time period into the prediction models of each modal component, calculate the predicted values of each modal component and add them up to obtain the dam deformation amount in the future time period.

[0011] Preferably, the S1 specifically includes the following steps:

[0012] S11. Collect the deformation monitoring data of the dam monitoring points and perform difference processing to obtain the initial continuous dam deformation data sequence;

[0013] S12. Collect the environmental quantity monitoring data and further process it at a certain time interval to obtain the continuous data sequence of the environmental quantity monitoring data; where the environmental quantity monitoring data includes historical reservoir water level monitoring data and reservoir area temperature monitoring data;

[0014] S13. Combine the initial continuous dam deformation data sequence, the continuous data sequence, and the corresponding time sequence to establish an initial sample data set.

[0015] Preferably, the expression of the initial sample data set is:

[0016]

[0017] Among them, represents the monthly average reservoir water level of the first observation month, represents the monthly average reservoir water level of the second observation month, represents the monthly average reservoir water level of the nth observation month, represents the monthly average temperature of the first observation month, represents the monthly average temperature of the second observation month, represents the monthly average temperature of the nth observation month, represents the monthly maximum temperature of the 1st observation month, represents the monthly maximum temperature of the 2nd observation month, represents the monthly maximum temperature of the nth observation month, represents the monthly minimum temperature of the 1st observation month, represents the monthly minimum temperature of the 2nd observation month, represents the monthly minimum temperature of the nth observation month, represents the average temperature of the current month and the previous month of the 1st observation month, represents the average temperature of the current month and the previous month of the 2nd observation month, represents the dam deformation of the 1st observation month, represents the dam deformation of the 2nd observation month, represents the dam deformation of the nth observation month.

[0018] Preferably, the S2 specifically includes the following steps:

[0019] S21. Input the initial continuous dam deformation data sequence into the variational mode decomposition algorithm and set the number of decomposition modes, and decompose the initial continuous dam deformation data sequence into several mode components;

[0020] S22. Merge the mode components with the continuous data sequence to obtain a mode sample data set, and the expression of the mode sample data set is specifically as follows:

[0021]

[0022] where, respectively represent the mode components corresponding to the observation months.

[0023] Preferably, in the S3, the Huber function is used as the loss function of the weak regressor to construct the Huber weighted loss function, and the specific expression is:

[0024]

[0025] where, a is the difference between the predicted value and the true value, and δ is a preset threshold.

[0026] Preferably, the S3 specifically includes the following steps:

[0027] S31. Introduce the Huber loss function into the AdaBoost model, assign initial weights to the samples in each initial sample data set based on the Adaboost model, and train the weak regressor;

[0028] S32. Calculate the influence effect of each environmental quantity monitoring data on the target variable through the Spearman correlation coefficient, sort the weight values of each feature variable according to their magnitudes, and select the feature variable with the largest weight value as the space for splitting point selection.

[0029] S33. Calculate the prediction error of the weak regressor based on the Huber weighted loss function; conduct a one-by-one evaluation and selection in the space of the feature variable with the largest weight value, and finally select the splitting point with the smallest loss function value for splitting.

[0030] S34. Repeat the above steps. Each iteration will train a new weak regressor until the predetermined number of iterations is reached or the performance of the Adaboost model meets the requirements.

[0031] Preferably, the S4 specifically includes the following steps:

[0032] S41. Divide the modal sample data set into a training set, a validation set, and a test set according to the ratio of 7:2:1.

[0033] S42. Set the types and search spaces of the hyperparameters of the DpAdaBoost model and establish the DpAdaBoost model.

[0034] S43. Use the GridSampler optimization algorithm to perform hyperparameter optimization on the DpAdaBoost model, set the number of optimization iteration training times and evaluation metrics, output the optimal hyperparameters and substitute them into the DpAdaBoost model.

[0035] S44. Use the training set and the validation set as inputs and each modal component as outputs, and train them respectively in the DpAdaBoost model to obtain the prediction models of each modal component with the optimal hyperparameters.

[0036] S45. Use the test set as the input, obtain the prediction results of each modal component through the prediction models of each modal component with the optimal hyperparameters, and calculate the evaluation metrics of the prediction results of each modal component and the actual modal components in the test set to verify the accuracy of each IMF component prediction model.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] 1. In this solution, by introducing the Huber function as the loss function of the weak regressor, compared with other loss functions, the Huber loss function is less sensitive to outliers or noise, which helps to reduce the influence of outliers on model training. When selecting the splitting point of the decision tree weak regressor each time, introducing a dynamic feature variable weight strategy can effectively reduce the computational amount of the splitting node selection step and greatly improve the model training efficiency.

[0039] 2. The DpAdaboost model of this solution makes targeted improvements to the dam deformation prediction problem through Step 3, and the DpAdaboost model has high prediction accuracy. Utilize the high optimization efficiency of the GridSampler algorithm to quickly select the optimal AdaBoost model hyperparameter combination. Description of the Drawings

[0040] Figure 1 Schematic diagram of the overall process of the embodiment of the present invention;

[0041] Figure 2 Comparison diagram of the continuous data sequences of each IMF component and the denoised Y in the embodiment of the present invention;

[0042] Figure 3 Comparison diagram of the predicted dam deformation amount and the actual dam deformation amount in the embodiment of the present invention. Detailed Embodiment

[0043] Embodiment 1: As Figure 1 - Figure 3 shown, a new method for improving the efficiency and accuracy of dam deformation prediction, as Figure 1 shown, is specifically implemented according to the following steps:

[0044] Step 1, obtain the historical continuous deformation monitoring data and continuous environmental quantity monitoring data of the dam monitoring points, and preprocess them to establish an initial sample data set.

[0045] Step 1.1, collect the historical continuous deformation monitoring data corresponding to the dam monitoring points. Since the monitoring frequencies in different periods are not equal, perform difference processing on them to uniformly process them into a continuous deformation data sequence with a monthly time interval, and obtain an initial dam continuous deformation data sequence Y, whose expression is:

[0046]

[0047] Among them, represents the dam deformation amount in the first observation month, represents the dam deformation amount in the second observation month, represents the dam deformation amount in the nth observation month.

[0048] Step 1.2, collect the historical continuous environmental quantity monitoring data of the dam, specifically including historical reservoir water level monitoring data and reservoir area temperature monitoring data, and further process them into H P , T P , T MAX , T MIN , T 60 , where H P represents the monthly average reservoir water level, T PDenote the monthly average temperature as \(T\). MAX Denote the monthly maximum temperature as \(T\). MIN Denote the monthly minimum temperature as \(T\). 60 Denote the average temperatures of the current month and the previous month, and obtain \(H\), P \(T\), P \(T\), MAX \(T\), MIN \(T\), 60 respectively, to form a continuous data sequence of \(T\).

[0049] Step 1.3: Merge the initial continuous deformation data sequence \(Y\) obtained in Step 1.1, the processed data of the continuous data sequences of \(H\), P \(T\), P \(T\), MAX \(T\), MIN \(T\), 60 \(T\) obtained in Step 1.2, and the corresponding time series \(\theta\), where \(\theta\) is the number of days of cumulative observation divided by 100, to form an initial sample data set. The expression of the initial sample data set is:

[0050]

[0051] where,[[]] denotes the monthly average reservoir water level of the first observation month, denotes the monthly average reservoir water level of the second observation month, denotes the monthly average reservoir water level of the \(n\)th observation month, and so on for others.

[0052] Step 2: Use the VMD (Variational Mode Decomposition) algorithm to decompose the deformation monitoring data in the initial sample data set obtained in Step 1 into several IMF components, and merge them with the data of \(H\), P \(T\), P \(T\), MAX \(T\), MIN \(T\), 60 \(T\), \(\theta\) obtained in Step 1 to establish an IMF sample data set.

[0053] Step 2.1: Input the initial continuous deformation data sequence \(Y\) obtained in Step 1.1 into the VMD algorithm, and set the number \(k\) of decomposition modes (IMF), and decompose \(Y\) into \(k\) IMF components.

[0054] Among them, when setting the value of k, if the value of k is set too small, it is easy to cause mode aliasing in the decomposition, that is, the frequencies of two or more IMF components are similar. When the value of k is large, it will lead to over-decomposition and generate useless components. Therefore, the selection of the value of k can be determined by observing whether there is aliasing in each IMF component after decomposition. For example, if k is set to 2, observe whether there is aliasing in the 2 IMF components after decomposition. If there is aliasing, modify k to 3 and continue to decompose and observe whether there is aliasing, and so on, until there is no aliasing in the k IMF components after decomposition. At this time, the value of k is the best.

[0055] Step 2.2: Combine the k IMF components obtained in Step 2.1 with the continuous data sequences of H P , T P , T MAX , T MIN , T 60 , and θ obtained from the initial sample dataset in Step 1.3 to obtain an IMF sample dataset. The specific expression of the dataset is as follows:

[0056]

[0057] Among them, respectively represent the k IMF components obtained in Step 2.1.

[0058] Step 3: Based on the Adaboost model, in the weak regressor node splitting step of the Adaboost model, considering the influence of each environmental quantity factor on the dam deformation effect, introduce a dynamic feature variable weight strategy, and use the Huber function as the loss function of the weak regressor to establish the DpAdaBoost model.

[0059] The main function of this step is to improve the Adaboost model according to the characteristics of the dam deformation prediction problem and establish the DpAdaBoost model. Here, a brief overview of the basic principle of the Adaboost model is given to facilitate the detailed description of the improvement strategy below. The basic principle of the Adaboost model is as follows:

[0060] First, initialize the sample weights, assign an initial weight to each sample, and usually these weights are equal; second, train the weak regressor, use the current sample weights to train a weak regressor, such as a decision tree; then, calculate the error and update the weights, calculate the prediction error of the weak regressor, and adjust the sample weights according to the error; next, resample, resample the data according to the updated sample weights. Finally, repeat the above steps. Each iteration will train a new weak regressor until the predetermined number of iterations is reached or the model performance meets the requirements.

[0061] For the training of the weak regressor, the present invention improves it with the decision tree as the weak regressor. Therefore, the principle when the weak regressor is a decision tree is described in detail here. The detailed principle of training the weak regressor in each round is as follows:

[0062] (1) Selection of splitting point: Find a splitting point in the feature space so that after splitting from this splitting point, the calculated loss function value is the smallest. Each split evaluates all features and all possible splitting points, that is, calculates the corresponding loss function value, and finally selects the splitting point with the smallest loss function value for splitting. The principle of splitting point selection is as follows:

[0063] a. Weighted loss function: In the AdaBoost model, the loss function is weighted, which takes into account the weights of each sample, and its calculation formula is modified to consider the sample weights. Taking the mean squared error (MSE) as an example, when selecting the splitting point, the weighted mean squared error (WMSE) will be calculated. The calculation of WMSE is as follows:

[0064]

[0065] where m represents the number of samples, w i is the weight of the i-th sample, y i is the actual target value, is the predicted value of the sample under the splitting point.

[0066] b. Considering the influence of weights: When calculating the WMSE of each possible splitting point, samples with higher weights will have a greater impact on the total loss. Therefore, when the decision tree is looking for the best splitting point, it will tend to those splits that can better reduce the weighted loss.

[0067] c. Select the best split: Finally, the decision tree will select the splitting point that can minimize the weighted loss to be the splitting node. This means that if the prediction errors of some samples in the current model are large (so the weights are high), then the decision tree will pay more attention to these samples when choosing the root node.

[0068] (2) Generate leaf nodes: When the stopping condition is reached, this node becomes a leaf node and is assigned a predicted value. This predicted value is usually the average or median of the sample target values on this node.

[0069] (3) Build a complete decision regression tree: Repeat steps 1 and 2 until all nodes become leaf nodes to build a complete decision regression tree.

[0070] Step 3.1, since there are often abnormal error points in the dam deformation monitoring data and environmental quantity monitoring data, which causes the Adaboost model to be affected by the abnormal error points during training. To address this issue, this step introduces the Huber function as the loss function of the weak regressor. Compared with other loss functions, the Huber loss function is less sensitive to outliers or noise. The principle is as follows:

[0071]

[0072] Among them, a is the difference between the predicted value and the true value (i.e., the residual), and δ is a preset threshold value, which is between 2 and 10. This threshold value will be optimized and selected as a hyperparameter during actual training. When the residual a is small (less than or equal to δ), the behavior of the Huber loss function is similar to MSE, that is, it squares small errors, which helps the model to learn precisely within the small error range. When the residual a is large (greater than δ), the behavior of the Huber loss function is similar to the absolute error loss, that is, it linearly processes the errors, which helps to reduce the impact of outliers on model training.

[0073] When the Huber loss function is introduced into the AdaBoost model, to consider the sample weights, the calculation formula of its Huber weighted loss function is modified as:

[0074]

[0075] Step 3.2, during the training of the weak regressor of the Adaboost model, when selecting the splitting point at each time, it is necessary to evaluate all features and all possible splitting points, calculate the corresponding loss function values, and finally select the splitting point with the smallest loss function value for splitting. In the dam deformation prediction problem, there are often many types of features, and the process of evaluating all features and all possible splitting points will consume a large amount of computing resources, resulting in low model training efficiency. To address this issue, this step introduces a dynamic feature variable weight strategy when selecting the splitting point of the decision tree weak regressor each time. The specific method steps are as follows:

[0076] Step 3.2.1, update the weight values of each feature variable of all samples on the current node. The greater the influence effect of the feature variable on the target variable, the higher the weight value. The derivation of the weight value formula is as follows:

[0077] The Spearman correlation coefficient is used to characterize the influence effect size of each feature variable on the target variable. The calculation method is as follows:

[0078]

[0079] where r is the Spearman correlation coefficient, d iis the rank difference between the feature variable and the target variable, and n is the number of samples at the current node.

[0080] The weight value formula of each feature variable of all samples at the current node is as follows:

[0081]

[0082] where α j represents the weight value of the j-th feature variable, r j represents the Spearman correlation coefficient of the j-th feature variable, and M represents the number of feature variables.

[0083] Step 3.2.2: According to the weight values of each feature variable of all samples at the current node obtained in Step 3.2.1, sort the weight values of each feature variable by size, select the feature variable with the largest weight value as the space for splitting point selection, and then use the Huber weighted loss function calculation formula obtained in Step 3.1 to evaluate and select one by one in the space of the feature variable with the largest weight value, and finally select the splitting point with the smallest loss function value for splitting.

[0084] Step 3.3: Use the methods described in Steps 3.1 and 3.2 respectively to improve the Adaboost model, and construct the DpAdaBoost model for the dam deformation prediction problem.

[0085] Step 4: Use H P , T P , T MAX , T MIN , T 60 , θ in the IMF sample dataset obtained in Step 2 as the input of the DpAdaBoost model, and each IMF component as the output respectively. And introduce the GridSampler optimization algorithm to automatically optimize the hyperparameters of the DpAdaBoost model, obtain the optimal hyperparameters, and thus obtain the prediction models of each IMF component with the optimal hyperparameters.

[0086] Step 4.1: Divide the IMF sample dataset obtained in Step 2 into a training set, a validation set, and a test set according to the ratio of 7:2:1.

[0087] Step 4.2: Set the types and search spaces of the hyperparameters of the DpAdaBoost model, establish the DpAdaBoost model, set the types of hyperparameters to be the maximum depth of the tree, the number of base estimators, the learning rate, and the threshold δ of the Huber loss function respectively, and set the value spaces of the hyperparameters to be 1 - 7, 50 - 200, 0.01 - 1, and 2 - 10 respectively.

[0088] Step 4.3: Establish a GridSampler-DpAdaBoost model. Use the GridSampler optimization algorithm to perform hyperparameter optimization on the DpAdaBoost model established in Step 4.2. Set the number of optimization iteration training times n_trials, and use R 2 as the evaluation index of the model training effect. After the optimization iteration training is completed, output the hyperparameter combination corresponding to the maximum R 2 during the n_trials training times.

[0089] Step 4.4: Use H P , T P , T MAX , T MIN , T 60 , and θ in the training set and validation set divided in Step 4.1 as inputs, and use each IMF component as the output. Train them respectively in the GridSampler-DpAdaBoost model established in Step 4.3 to obtain the prediction models of each IMF component with optimal hyperparameters.

[0090] Step 4.5: Input H P , T P , T MAX , T MIN , T 60 , and θ in the test set divided in Step 4.1 into the prediction models of each IMF component with optimal hyperparameters obtained in Step 4.4 to obtain the prediction results of each IMF component. Then, compare and calculate the R 2 between the prediction results of each IMF component and the actual IMF components in the test set divided in Step 4.1 to verify the accuracy of each IMF component prediction model and ensure the prediction effect of the model.

[0091] Step 5: Input the H P , T P , T MAX , T MIN , T 60 , and θ values corresponding to a certain future month or several months (here may be predicted values, and the environmental factors can be estimated according to experience based on the month) into the prediction models of each IMF component obtained in Step 4, output the predicted values of each IMF component, and add and combine the predicted values of each IMF component to obtain the dam deformation amount in a certain future month or several months.

[0092] In a new method for improving the efficiency and accuracy of dam deformation prediction in the present invention: The function of Step 2 is to decompose the deformation monitoring data through the VMD algorithm, thereby reducing the complexity of prediction.

[0093] The VMD algorithm is adopted, and its principle is as follows: The decomposition process of VMD is the process of solving a variational problem. In this algorithm, the intrinsic mode function (IMF) is defined as an amplitude-modulated and frequency-modulated function with bandwidth limitation. The function of the VMD algorithm is to decompose the original signal into a specified number of IMF components by constructing and solving a constrained variational problem.

[0094] The advantages of adopting the VMD algorithm are as follows: VMD transfers the process of obtaining signal components into a variational framework and adopts a non-recursive processing strategy. By constructing and solving a constrained variational problem, the decomposition of the original signal is realized, which can effectively avoid problems such as mode mixing, over-envelope, under-envelope, and boundary effects, and has advantages such as better complex data decomposition accuracy and better anti-noise interference.

[0095] In a novel method for improving the efficiency and accuracy of dam deformation prediction in the present invention: The function of step 3 is to improve the Adaboost model according to the characteristics of the dam deformation monitoring problem and establish a DpAdaBoost model.

[0096] The principle of establishing the DpAdaBoost model is as follows: Based on the Adaboost model, according to the characteristics of the dam deformation monitoring problem, in the weak regressor node splitting step of the Adaboost model, considering the influence degree of each environmental quantity factor on the dam deformation effect, a dynamic feature variable weight strategy is introduced to optimize the calculation method of the weak regressor node splitting step; and the Huber function is introduced as the loss function of the weak regressor to establish the DpAdaBoost model.

[0097] The advantages of adopting the DpAdaBoost model are as follows: The Huber function is introduced as the loss function of the weak regressor. Compared with other loss functions, the Huber loss function is less sensitive to outliers or noise, which helps to reduce the influence of outliers on model training. When selecting the splitting point of the decision tree weak regressor each time, introducing a dynamic feature variable weight strategy can effectively reduce the calculation amount of the splitting node selection step and greatly improve the model training efficiency.

[0098] In a novel method for improving the efficiency and accuracy of dam deformation prediction in the present invention: The function of step 4 is to establish a GridSampler-DpAdaboost model, use the GridSampler algorithm to optimize the hyperparameters of the DpAdaboost model, and finally obtain a prediction model of each IMF component with optimal hyperparameters through training and learning.

[0099] The GridSampler-DpAdaboost model is adopted, and its principle is as follows: The GridSampler optimization algorithm is used to optimize the hyperparameters of the DpAdaboost model during the training process, quickly select the optimal combination of DpAdaboost model hyperparameters, and thus establish the optimal prediction model for each IMF component.

[0100] The advantages of the GridSampler-DpAdaboost model are as follows: Through step 3, the DpAdaboost model makes targeted improvements to the dam deformation prediction problem, and the DpAdaboost model has high prediction accuracy. Utilizing the high optimization efficiency of the GridSampler algorithm, quickly select the optimal combination of AdaBoost model hyperparameters.

[0101] Example 2: The following is an example for a detailed description of the present invention.

[0102] 1. Obtain the historical continuous deformation monitoring data and historical continuous environmental quantity monitoring data of a certain monitoring point on the dam crest of a certain dam from January 2014 to December 2023, and perform preprocessing according to the method described in step 1 to obtain a total of 120 groups of data, and construct an initial sample data set.

[0103] 2. According to the method described in step 2, decompose the deformation monitoring data in the initial sample data set into 3 IMF components by using the VMD (Variational Mode Decomposition) algorithm, and merge them with the H P , T P , T MAX , T MIN , T 60 , θ data obtained in step 1 to obtain an IMF sample data set. The comparison diagrams of the continuous data sequences of each IMF component and the denoised Y are as Figure 2 shown.

[0104] 3. Based on the Adaboost model, in the weak regressor node splitting step of the Adaboost model, considering the influence of each environmental quantity factor on the dam deformation effect, introduce a dynamic feature variable weight strategy, and use the Huber function as the loss function of the weak regressor to establish a DpAdaBoost model.

[0105] 4. According to the method described in step 4, divide the IMF sample data set into a training set, a validation set, and a test set according to the ratio of 7:2:1. The number of samples in each data set is 84 groups, 24 groups, and 12 groups respectively. Set the optimized training iteration times of the GridSampler-DpAdaBoost model to 100 times, input the training set and the validation set into the GridSampler-DpAdaBoost model for training, and obtain the prediction models of each IMF component with the optimal hyperparameters respectively.

[0106] 5. Input H P , T P , T MAX , T MIN , T 60 , and θ into each IMF component prediction model respectively to obtain the prediction results of each IMF component prediction. Merge them to obtain the predicted dam deformation amount. Compare and calculate the R 2 with the actual dam deformation amount after merging each IMF component in the validation set, which is 0.823. The comparison graph is as Figure 3 shown, and the prediction accuracy is good.

Claims

1. A novel method for improving the efficiency and accuracy of dam deformation prediction, characterized in that, It includes the following steps: S1. Obtain the deformation monitoring data and environmental quantity monitoring data of the dam monitoring points, perform preprocessing, and establish an initial sample data set; S2. Based on the initial sample data set, use the variational mode decomposition algorithm to decompose it into several modal components, and combine with the environmental quantity monitoring data to establish a modal sample data set; S3. Based on the Adaboost model, introduce a dynamic feature variable weight strategy when splitting the weak regressor nodes of the Adaboost model and use the Huber function as the loss function of the weak regressor to establish a DpAdaBoost model; S4. Use the modal sample data set as the input, each modal component as the output, and introduce the GridSampler optimization algorithm to automatically optimize the hyperparameters of the DpAdaBoost model to obtain the optimal hyperparameters, so as to obtain the prediction models of each modal component with the optimal hyperparameters; S5. Input the environmental quantity monitoring data corresponding to the future time period into the prediction models of each modal component, calculate the predicted values of each modal component and add them up to obtain the dam deformation quantity in the future time period.

2. A novel method for improving the efficiency and accuracy of dam deformation prediction according to claim 1, characterized in that, The specific steps of S1 are as follows: S11. Collect the deformation monitoring data of the dam monitoring points and perform difference processing to obtain an initial continuous dam deformation data sequence; S12. Collect the environmental quantity monitoring data and further process it at a certain time interval to obtain a continuous data sequence of the environmental quantity monitoring data; where the environmental quantity monitoring data includes historical reservoir water level monitoring data and reservoir area temperature monitoring data; S13. Combine the initial continuous dam deformation data sequence, the continuous data sequence and the corresponding time sequence to establish an initial sample data set.

3. A novel method for improving the efficiency and accuracy of dam deformation prediction according to claim 2, characterized in that, The expression of the initial sample data set is: Among them, represents the monthly average reservoir water level of the 1st observation month, represents the monthly average reservoir water level of the 2nd observation month, represents the monthly average reservoir water level of the nth observation month, represents the monthly average temperature of the 1st observation month, represents the monthly average temperature of the 2nd observation month, represents the monthly average temperature of the nth observation month, represents the monthly maximum temperature of the 1st observation month, represents the monthly maximum temperature of the 2nd observation month, represents the monthly maximum temperature of the nth observation month, represents the monthly minimum temperature of the 1st observation month, represents the monthly minimum temperature of the 2nd observation month, represents the monthly minimum temperature of the nth observation month, represents the average temperature of the current month and the previous month of the 1st observation month, represents the average temperature of the current month and the previous month of the 2nd observation month, represents the dam deformation of the 1st observation month, represents the dam deformation of the 2nd observation month, represents the dam deformation of the nth observation month.

4. A novel method for improving the efficiency and accuracy of dam deformation prediction according to claim 1, characterized in that, The specific steps of S2 are as follows: S21. Input the initial continuous dam deformation data sequence into the variational mode decomposition algorithm and set the number of decomposition modes to decompose the initial continuous dam deformation data sequence into several modal components; S22. Combine the modal components with the continuous data sequence to obtain a modal sample data set. The specific expression of the modal sample data set is as follows: Among them, respectively represent the modal components corresponding to the observed months.

5. A novel method for improving the efficiency and accuracy of dam deformation prediction according to claim 1, characterized in that, In S3, use the Huber function as the loss function of the weak regressor to construct a Huber weighted loss function. The specific expression is: Among them, a is the difference between the predicted value and the true value, and δ is a preset threshold.

6. A novel method for improving the efficiency and accuracy of dam deformation prediction according to claim 5, characterized in that, The specific steps of S3 are as follows: S31. Introduce the Huber loss function into the AdaBoost model, assign initial weights to the samples in each initial sample data set based on the Adaboost model, and train the weak regressor; S32. Calculate the influence effect of each environmental quantity monitoring data on the target variable through the Spearman correlation coefficient, sort the weight values of each feature variable according to the size, and select the feature variable with the largest weight value as the space for splitting point selection; S33. Calculate the prediction error of the weak regressor based on the Huber weighted loss function; perform a one-by-one evaluation and selection in the feature variable space with the largest weight value, and finally select the splitting point with the smallest loss function value for splitting; S34. Repeat the above steps. A new weak regressor will be trained in each iteration until the predetermined number of iterations is reached or the performance of the Adaboost model meets the requirements.

7. A novel method for improving the efficiency and accuracy of dam deformation prediction according to claim 1, characterized in that, The specific steps of S4 are as follows: S41. Divide the modal sample dataset into a training set, a validation set, and a test set according to the ratio of 7:2:

1. S42. Set the types and search spaces of the hyperparameters of the DpAdaBoost model and establish the DpAdaBoost model. S43. Use the GridSampler optimization algorithm to optimize the hyperparameters of the DpAdaBoost model, set the number of optimization iteration training times and evaluation metrics, output the optimal hyperparameters and substitute them into the DpAdaBoost model. S44. Take the training set and the validation set as inputs and the modal components as outputs, and train them in the DpAdaBoost model respectively to obtain the prediction models of each modal component with the optimal hyperparameters. S45. Take the test set as the input, obtain the prediction results of each modal component through the prediction models of each modal component with the optimal hyperparameters, and calculate the evaluation metrics of the prediction results of each modal component and the actual modal components in the test set to verify the accuracy of each IMF component prediction model.

Images