Dam seepage intelligent prediction method introducing key sample learning strategy

Abstract

The invention discloses a dam seepage intelligent prediction method introducing a key sample learning strategy, and belongs to the technical field of seepage prediction, and the method comprises the steps: building a novel dam seepage intelligent prediction TPESampler-IKSLS-AdaBoost model introducing the key sample learning strategy, and enabling the IKSLS-AdaBoost model to introduce the key sample learning strategy into the TPESampler-IKSLS-AdaBoost model, due to the fact that dam external environment quantity effects in similar months in different years are similar, dam seepage effects are also similar, and therefore when the to-be-predicted accumulated seepage flow of a target dam in a certain period of time in the future is predicted, the to-be-predicted accumulated seepage flow is predicted. Weights of samples which are in a sample data set and are similar to observation days of to-be-predicted accumulated seepage flow of a target dam in a certain period of time in the future in months are improved, key learning training is carried out, and the seepage prediction precision can be greatly improved. According to the method, the optimal AdaBoost model hyper-parameter combination is quickly selected by using the good optimization ability of the TPESampler algorithm in the aspect of nonlinear problems and the high optimization efficiency based on the KD tree structure, so that the prediction ability of the AdaBoost model is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of seepage prediction, and specifically relates to an intelligent dam seepage prediction method introducing a key sample learning strategy. Background Art

[0002] In water conservancy projects, objects such as reservoir dams are key construction facilities, playing a core role in aspects such as flood control, agricultural irrigation, water transportation, and power production. During the operation process of objects such as reservoir dams after completion, the seepage effect is one of the main factors affecting the safety of objects such as reservoir dams. The seepage state can directly reflect the working state of the object and plays an important role in the safety monitoring of the object. Currently, the methods and equipment for seepage monitoring of objects have been very mature, and automated real-time monitoring of object seepage can be achieved. However, the development of seepage prediction for objects such as reservoir dams lags behind. Precise seepage prediction of objects can detect potential seepage problems in advance, timely detect changes in the safety situation of the object, and ensure the safe operation of the object.

[0003] Dam seepage prediction is crucial for ensuring the safe operation of hydraulic structures. Accurate seepage prediction helps prevent dam-related accidents and safeguard the safety of personnel and property. Dam seepage is a complex physical process, jointly affected by multiple factors such as geological structure, hydrogeological properties, the design and material characteristics of the dam itself, water pressure, and environmental temperature. Traditional prediction methods include prediction based on empirical formulas and numerical model analysis. These methods have their limitations. For example, empirical formulas overly simplify the problem, the numerical model calculation process is complex, requires a large amount of data, and often simplifies the actual situation with assumptions to reduce the calculation amount, resulting in unsatisfactory prediction accuracy and efficiency.

[0004] With the development of machine learning technology, it provides new ideas for solving the problem of dam seepage prediction. The AdaBoost model, as a powerful ensemble learning technology, can achieve a very high prediction accuracy by gradually training and boosting weak learners. However, how to use the AdaBoost model for dam seepage prediction, effectively select the hyperparameter combination of the AdaBoost model, and improve the AdaBoost model for the dam seepage prediction problem to optimize the performance of the AdaBoost model in dam seepage prediction is still a challenge.

[0005] The Chinese patent, CN118333433A, proposes a target object seepage prediction method based on the BO-GBRT strategy. It uses the machine learning algorithm of gradient boosting regression tree to construct a seepage prediction model, and introduces the Bayesian optimization algorithm to optimize the hyperparameters of the gradient boosting regression tree model, solving the problem of complex optimization and tuning of the gradient boosting regression tree. However, this application also has problems such as unreasonable data screening, cumbersome and complex hyperparameter optimization process, and the training process will contaminate the model. Summary of the Invention

[0006] The object of the present invention is to address the problems in the prior art for dam seepage prediction, such as unreasonable data screening, cumbersome and complex hyperparameter optimization process, and the training process will contaminate the model. A dam seepage intelligent prediction method introducing a key sample learning strategy is proposed. By comprehensively considering and analyzing the factors affecting the seepage of the target object, based on the historical monitoring data of the target object seepage, the influencing factors are optimized by the convolutional smoothing filtering method, and the main factors affecting the target object seepage are screened out. The TPESampler optimization algorithm is used to optimize the hyperparameters of the AdaBoost model during the training process. Compared with the prior art, for the dam seepage prediction problem, a key sample learning strategy is specifically introduced, and the good optimization ability of the TPESampler algorithm for nonlinear problems and the high optimization efficiency based on the KD tree structure are utilized to quickly select the optimal AdaBoost model hyperparameter combination, thereby improving the prediction ability of the AdaBoost model. Through experimental verification, the prediction accuracy of the present invention is good, providing an effective solution for dam seepage prediction. Utilizing the good optimization ability of the TPESampler algorithm for multi-variable nonlinear problems and the high optimization efficiency based on the KD tree structure, quickly select the optimal AdaBoost model hyperparameter combination, thereby establishing the optimal seepage prediction model for the target dam and realizing the efficient and high-precision prediction of the seepage of the target dam.

[0007] To solve the above technical problems, the present invention provides a dam seepage intelligent prediction method introducing a key sample learning strategy, including the following steps:

[0008] S1. Obtain the data obtained by the dam seepage monitoring equipment and the environmental quantity monitoring equipment and perform preliminary processing, and then perform noise reduction processing using the convolutional smoothing filtering method to construct a monitoring data set;

[0009] S2. Based on the monitoring data set, divide the monitoring data set into a training set, a validation set, and a test set in a ratio of 7:2:1;

[0010] S3. Establish an IKSLS-AdaBoost model and perform iterative training on a weak regressor to obtain a new IKSLS-AdaBoost model, which is then trained using the training set and validation set as inputs to obtain the optimal hyperparameter combination for the dam seepage prediction model. Based on the optimal hyperparameter combination, a TPESampler-IKSLS-AdaBoost model is established.

[0011] S4. Use the test set to verify the TPESampler-IKSLS-AdaBoost model and perform seepage prediction.

[0012] Preferably, the S1 includes:

[0013] S11. Analyze the factors affecting seepage for specific engineering projects and obtain historical seepage monitoring data through corresponding monitoring equipment;

[0014] S12, based on the hysteresis effect, performing difference processing on the historical seepage monitoring data on a daily basis to obtain a continuous data sequence with daily time intervals;

[0015] S13, dividing the continuous data sequence into intervals of a user-defined observation time period to obtain a divided data sequence, wherein each divided data in the divided data sequence represents historical seepage monitoring data within the user-defined observation time period;

[0016] S14, integrating the historical seepage monitoring data corresponding to each partitioned data, combining the partitioned data sequence, and constructing an initial sample data set;

[0017] S15, using a convolution smoothing filter method, inputting the divided data into a convolution smoothing filter model respectively to perform noise reduction processing to obtain noise-reduced sample data;

[0018] S16. Based on the historical seepage monitoring data of the initial sample data set and the noise reduction sample data, a monitoring data set is constructed and normalized.

[0019] Preferably, the initial sample data set is represented as follows:

[0020]

[0021] in, C i ,lnC i is the characteristic value, which represents the initial target seepage influencing factor data; Y i is the target variable, representing the preferred sample data; N represents the number of factors affecting the initial target seepage; M represents the number of data groups in the sample data set.

[0022] Preferably, the specific expression for normalization is as follows:

[0023] X’ = (X - X min ) / (X max - X min );

[0024] where X’ is the value after normalization, X is the original data, X max is the maximum value in the dataset, and X min is the minimum value in the dataset.

[0025] Preferably, the specific expressions for the training set, validation set, and test set in S2 are:

[0026]

[0027] Validation set:

[0028]

[0029] Test set:

[0030]

[0031] Preferably, S3 includes the following steps:

[0032] S31. Based on the training set D, the number of samples in the dataset is m. Initialize the weights, that is, the sample distribution of the training set in the initial state is Dt1. The weight of each sample X i is initialized to 1 / m for training the first weak regressor h1, and Dt t is used for training the first weak regressor ht, and so on;

[0033] S32. Perform T iterations in a loop. In each iteration, the weak learner number is t, t ∈ {1, 2, 3,..., T}. This step is a loop body;

[0034] S321. On the basis of the sample weight distribution Dt t (X), train the weak classifier h t on the training set D; calculate the maximum error E t of the regressor h t on the training set D; calculate the squared error e t of h t for each sample according to E ti , so as to obtain the error rate e t of the current weak regressor h t , and finally calculate the weight w t of the weak regressor h t ;

[0035] S322. Update the weight distribution of the training set samples. The calculation formula for the maximum weight of the training set samples in the next iteration training is expressed as follows:

[0036]

[0037] where Dt t+1 (X)max represents the maximum weight of the dataset samples in the next iteration training;

[0038] S323. Update the weights of all key samples in the next iteration training to Dt t+1 (X)max, that is, the calculation formula for the weights of other samples is as follows:

[0039]

[0040] where Z t is the normalization factor, and its calculation formula is: If X i is a key sample, then is Dt t+1 (X)max, otherwise is

[0041] S324. Let t = t + 1, and return to S321 to continue the iteration. After T rounds of iteration, the IKSLS-AdaBoost model is finally obtained;

[0042] S33. Establish a function. Each time this function is called, the AdaBoost model corresponding to different hyperparameter combinations will be trained once and the training effect R 2 ;

[0043] S331. Establish a function. First, set the types and search spaces of the hyperparameters of the IKSLS-AdaBoost model in the function;

[0044] S332. Based on the value-taking rule of the TPESampler optimization algorithm, each time the function is called, a set of hyperparameter values of the IKSLS-AdaBoost model is taken, and a new IKSLS-AdaBoost model is established;

[0045] S333. Using the training set as the input, train based on the new IKSLS-AdaBoost model, then verify the training effect using the validation set, and calculate and return the R 2 ;

[0046] S34. Call the established function, set the number of training times n_trials, and set the optimal hyperparameters to be the optimal hyperparameter combination in n_trials training times, thus completing the establishment of the TPESampler-KSLS-AdaBoost model.

[0047] Preferably, the obtaining of the hyperparameter combination corresponding to the optimal training effect in n_trials training times includes calculating the R of the predicted data of the validation set 2 and comparing it with a preset threshold. If the R of the validation set 2 is greater than or equal to the preset threshold, output the hyperparameter combination. If the R of the predicted data 2 is less than the preset threshold, continue training through the training set within the set number of training times.

[0048] Preferably, the types of hyperparameters include the number of decision trees, learning rate, maximum depth of the tree, minimum number of samples required to split a node, and minimum number of samples that a leaf node should have.

[0049] Preferably, S4 includes: inputting the test set into the optimal seepage prediction model for verification. When predicting the cumulative seepage volume of the target dam in the future time period, taking the first day as the observation day, obtaining the historical seepage volume monitoring data corresponding to the seepage volume on the observation day through the dam seepage monitoring equipment and environmental quantity monitoring equipment, performing normalization processing, and then conducting seepage volume prediction.

[0050] Advantages of the present invention:

[0051] 1. This solution uses the AdaBoost model to train the historical seepage monitoring data and environmental quantity data of the dam, and can achieve the prediction of the future seepage situation of the dam. In addition, the TPESampler optimization algorithm is used to optimize the hyperparameters of the AdaBoost model during the training process. Utilizing the good optimization ability of the TPESampler algorithm for multi-variable non-linear problems and the high optimization efficiency based on the KD-tree structure, quickly select the optimal AdaBoost model hyperparameter combination, thereby establishing the optimal seepage prediction model for the target dam, realizing the efficient and high-precision prediction of the seepage of the target dam, discovering dam seepage anomalies in advance, supporting the daily operation management of the dam, and ensuring the healthy and safe operation of the dam.

[0052] 2. This solution reduces random noise and smooths the signal by sliding a small window on the data sequence and performing weighted averaging on the data within the window. When denoising the data sequence, this method can effectively remove noise interference while retaining the inherent trend and characteristics of the signal.

[0053] 3. In this solution, the IKSLS-AdaBoost model is adopted to introduce the key sample learning strategy. Since the external environmental quantities of the dam in similar months of different years have similar effects, resulting in similar seepage effects of the dam, when predicting the cumulative seepage volume to be predicted in a certain period in the future of the target dam, the weights of the samples in the sample dataset whose observation dates of the cumulative seepage volume to be predicted in a certain period in the future of the target dam are similar in months are increased for key learning and training, which can greatly improve the accuracy of seepage prediction. At the same time, the underlying logic of the AdaBoost model is changed, and the good optimization ability of the TPESampler algorithm in nonlinear problems and the high optimization efficiency based on the KD tree structure are used to quickly select the optimal AdaBoost model hyperparameter combination, thereby improving the prediction ability of the AdaBoost model. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flowchart of an embodiment of an intelligent dam seepage prediction method introducing a key sample learning strategy according to the present invention;

[0055] Figure 2 It is a comparison curve graph of the example prediction value and the actual value of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] Embodiment 1: As Figure 1 and Figure 2 shown, an intelligent dam seepage prediction method introducing a key sample learning strategy includes the following steps:

[0057] Step 1, obtain the historical monitoring data of the earth-rock dam seepage and the influencing factor data of the earth-rock dam seepage, and process the obtained data to construct an initial sample dataset.

[0058] Step 1.1, the main influencing factors affecting the earth-rock dam seepage include temperature factor, rainfall factor, water pressure factor and aging factor. Therefore, the data required for earth-rock dam seepage prediction include temperature, rainfall, water pressure, historical monitoring data of earth-rock dam seepage and aging factor data. Among them, the temperature, rainfall, water pressure and historical monitoring data of earth-rock dam seepage are obtained through corresponding monitoring devices. The temperature is the temperature of the reservoir area, the rainfall is the rainfall data of the reservoir area, the water pressure is the reservoir water level, and the aging factor data is the cumulative number of days since the earth-rock dam was built and put into operation.

[0059] Step 1.2, since the monitoring frequencies of various monitoring devices of the earth-rock dam are different in different periods, therefore, the difference processing is respectively carried out on the collected reservoir area temperature, reservoir area rainfall, reservoir water level and historical monitoring data of earth-rock dam seepage in step 1.1 on a daily basis, and they are uniformly processed into a continuous data sequence with a daily time interval.

[0060] Step 1.3. Since the change in the seepage flow rate of an earth-rock dam always lags behind the changes in the reservoir area temperature, reservoir area rainfall, and reservoir water level, and the lag effects of different influencing factors on the earth-rock dam vary. To further determine which time period's influencing factor values before the seepage prediction date play a major role in affecting the seepage flow rate on the seepage prediction date, based on the magnitudes of the lag effects of different influencing factors, the determined initial target seepage influencing factors are divided into several influencing factors according to time periods. The present invention further divides the three influencing factors of reservoir area temperature, reservoir area rainfall, and reservoir water level into P Q1-10 、L R1-2 、L R2-3 、P H1-3 、P H3-5 、P H5-10 、P H10-15 a total of 7 factors. Among them, P Q1-10 represents the average reservoir area temperature 1 - 10 days before the seepage observation date, L R1-2 represents the cumulative rainfall 1 - 2 days before the seepage observation date, P H1-3 represents the average reservoir water level 1 - 3 days before the seepage observation date, and so on. In addition, the aging influencing factor is divided into C and lnC, where C represents the number of days the earth-rock dam has been in operation on the seepage observation date. Therefore, the present invention divides the factors affecting the seepage of the earth-rock dam into P Q1-10 、L R1-2 、L R2-3 、P H1-3 、P H3-5 、P H5-10 、P H10-15 、C and lnC, a total of 9.

[0061] It can be understood that the seepage flow rates at different positions of the dam are different, and the influencing factors are also different. Using the historical seepage flow rate monitoring data at a certain position of the dam to construct sample data can more precisely analyze the change of the seepage flow rate of the dam. At the same time, the sample data can be the seepage flow rate monitoring value at a certain position or the set of seepage flow rate monitoring values at multiple positions.

[0062] Step 1.4. The processed historical seepage monitoring data of the earth-rock dam in Step 1.2 is further divided into data sequences at 3-day time intervals, where each data represents the cumulative seepage flow rate of the earth-rock dam in 3 days, and the seepage observation date of the corresponding influencing factor is based on the first day of the 3 days.

[0063] Step 1.5. The processed earth-rock dam seepage data sequence obtained in Step 1.4 and the 9 influencing factors described in Step 1.3 corresponding to each seepage flow rate data are sorted and combined to construct an initial sample data set. The expression of the data set is as follows:

[0064]

[0065] Among them, Y is the target variable, which represents the seepage data of the earth-rock dam, and Y1 represents one of the seepage values; C 1 ,lnC 1 is the eigenvalue, which represents the seepage influencing factor corresponding to Y1, and so on. N represents the number of data groups in the sample data set.

[0066] Step 2: Since monitoring data often have abnormal error values, the sample data set obtained in step 1 is P Q1-10 、L R1-2 、L R2-3 、P H1-3 、P H3-5 、P H5-10 、P H10-15 And Y uses convolution smoothing filtering method to perform noise reduction to obtain a new sample data set.

[0067] Step 2.1: Use convolution smoothing filter to concentrate the initial sample data set obtained in step 1 into P Q1-10 、L R1-2 、L R2-3 、P H1-3 、P H3-5 、P H5-10 、P H10-15 and Y are input into the convolution smoothing filter model for noise reduction, and the noise-reduced PQ1-10, LR1-2, LR2-3, PH1-3, PH3-5, PH5-10, PH10-15 and Y are obtained respectively. The basic introduction of the convolution smoothing filter method is as follows:

[0068] Convolution filtering is a digital signal processing technique that reduces random noise and smoothes the signal by sliding a small window over the data sequence and performing a weighted average of the data within the window. This method effectively removes noise interference while preserving the signal's inherent trends and characteristics.

[0069] Step 2.2: The noise-reduced P obtained in step 2.1 is Q1-10 、L R1-2 、L R2-3 、P H1-3 、P H3-5 、P H5-10 、P H10-15 And Y replaces the corresponding data in the initial sample data set obtained in step 1 to form a new sample data set.

[0070] Step 3: Standardize each type of data in the new sample data set obtained in step 2, and obtain a standardized sample data set.

[0071] Among them, the expression for normalization processing is:

[0072] X' = (X - X min ) / (X max - X min );

[0073] Among them, X' is the value after normalization, X is the original data, X max is the maximum value in the dataset, and X min is the minimum value in the dataset.

[0074] Step 4: Divide the standardized sample dataset obtained in Step 3 into three parts, namely the training set, the validation set, and the test set, with a ratio of 7:2:1. Among them, the training set and the validation set are cross-validated during model training to improve the training accuracy, and the test set is used for the final evaluation of the model effect.

[0075] Divide the standardized sample dataset obtained in Step 3 into the training set, the validation set, and the test set according to the ratio of 7:2:1. The expressions for the three obtained datasets are as follows:

[0076] Training set:

[0077]

[0078] Validation set:

[0079]

[0080] Test set:

[0081]

[0082] Step 5: Establish a new type of intelligent prediction model for dam seepage, TPESampler-IKSLS-AdaBoost, which introduces the key sample learning strategy.

[0083] Step 5.1: Establish the IKSLS-AdaBoost model. That is, based on the AdaBoost machine learning algorithm process, when iteratively updating the sample weights, in view of the unique characteristics of the dam seepage prediction problem, introduce the key sample learning strategy specifically to establish the IKSLS-AdaBoost model. The following details the AdaBoost machine learning algorithm process and elaborates on the introduction principle and method of the key sample learning strategy.

[0084] Step 5.1.1: Assume there is a dataset D, which is also the training set. The number of samples in the dataset is m, and initialize the weights. Denote the sample distribution of the dataset in the initial state as Dt1. For each sample X i the weight is initialized to 1 / m, then Dt1(X i) = 1 / m. Dt1 is used for the training of the first weak regressor h1, and Dt t is used for the training of the first weak regressor h t and so on.

[0085] Step 5.1.2, perform T iterations in a loop. In each iteration, the weak learner number is t, where t ∈ {1, 2, 3, …, T}. This step is a loop body, and the steps in the loop body are as follows:

[0086] Step 5.1.2.1, based on the sample weight distribution Dt t (X), train the weak classifier h t on the dataset D; calculate the maximum error E t of the regressor h t on the training set D; according to E t calculate the squared error e t of h ti for each sample, so as to obtain the error rate e t of the current weak regressor h t , and finally calculate the weight w t of the weak regressor h t . The content described in this step is the principle of the well-known AdaBoost machine learning algorithm, so it is briefly described here.

[0087] Step 5.1.2.2, update the sample weight distribution of the dataset. Here, aiming at the unique characteristics of the dam seepage prediction problem, a key sample learning strategy is introduced to update the sample weights of the dataset.

[0088] In the AdaBoost machine learning algorithm, the greater the weight of a sample, the higher the attention paid to this sample during the iterative training of the subsequent weak regressors, that is, it will be focused on and learned intensively. Based on this feature, combined with the unique characteristics of the dam seepage prediction problem, a key sample learning strategy is introduced. Since the external environmental quantities of the dam are similar in similar months in different years, resulting in similar seepage effects of the dam, when predicting the cumulative seepage volume to be predicted for a certain period of time in the future of the target dam, the weights of the samples in the sample dataset whose observation dates of the cumulative seepage volume to be predicted for a certain period of time in the future of the target dam are close in month are increased for key learning and training, which can greatly improve the accuracy of seepage prediction.

[0089] (1) First, according to the observation date of the cumulative seepage volume to be predicted for a certain period of time in the future of the target dam, select all samples in the dataset whose months are the same as or one month different from this date (here, the difference in years is not considered), that is, the selected samples are key samples.

[0090] (2) Secondly, introduce the key sample learning strategy, that is, the key samples selected in (1) are used to update the current weak regressor ht When calculating the weights of samples, the highest sample weight is always maintained to ensure that key samples are given priority attention during the next iteration of the weak regressor training. To maintain the adaptive training effect of the AdaBoost machine learning model on other samples, a dynamic maximum weight method is used here to update the weight distribution of the dataset samples. The calculation process and formula derivation are as follows:

[0091] Formula for calculating the maximum weight of dataset samples during the next iteration of training:

[0092]

[0093] Dt t+1 (X) max The maximum weight of dataset samples during the next iteration of training.

[0094] Update the weights of all key samples during the next iteration of training to Dt t+1 (X) max , that is, the formula for calculating the weights of other samples is as follows:

[0095]

[0096] Among them, Z t is the normalization factor, and its calculation formula is:

[0097] If X i is a key sample, then is Dt t+1 (X) max , otherwise is

[0098] Step 5.1.2.3, let t = t + 1, and return to Step 5.1.2.1 to continue the iteration. After T rounds of iteration, the IKSLS-AdaBoost model is finally obtained.

[0099] Step 5.2, couple the TPESampler optimization algorithm with the IKSLS-AdaBoost model to establish the TPESampler-IKSLS-AdaBoost model.

[0100] Step 5.2.1, establish a function. Each time this function is called, the AdaBoost model corresponding to different combinations of hyperparameters will be trained once and the training effect R 2 will be returned, as follows:

[0101] Step 5.2.1.1, establish a function. First, set the types and search spaces of the hyperparameters of the IKSLS-AdaBoost model in the function.

[0102] Step 5.2.1.2. After setting the types and value spaces of the hyperparameters in Step 5.2.1.1, based on the value-taking rules of the TPESampler optimization algorithm, each time the function is called, a set of hyperparameter values of the IKSLS-AdaBoost model is taken, and a new IKSLS-AdaBoost model is established.

[0103] Step 5.2.1.3. Input the training set divided in Step 4 into the newly established IKSLS-AdaBoost model in Step 5.2.1.2 for training, then use the validation set to verify the training effect, calculate and return the R of the predicted data and the actual data of the validation set 2 。

[0104] It can be understood that the evaluation metrics can use a scoring function, including R 2 , accuracy, precision, recall, MSE (mean squared error), MAE (mean absolute error), CC / PCC (correlation coefficient), etc. Among them, R 2 is the evaluation metric we usually use. R 2 is used as an evaluation metric to evaluate the fitting degree of the model to the data. Its value range is from 0 to 1, and the closer it is to 1, the better the fitting degree of the model. In statistics, the R 2 evaluation metric usually represents the accuracy of the model's prediction. The closer the value is to 1, the better the prediction effect of the model. On the contrary, if the value is close to 0, it indicates that the prediction effect of the model is worse.

[0105] Step 5.2.2. Call the function established in Step 5.2.1, set the number of training times n_trials, and set the optimal hyperparameters to take the hyperparameter combination corresponding to the optimal training effect in n_trials training. Through the above steps, the establishment of the TPESampler-IKSLS-AdaBoost model is completed.

[0106] Obtaining the hyperparameter combination corresponding to the optimal training effect in n_trials training includes calculating the R of the predicted data of the validation set 2 and comparing it with the preset threshold. If the R of the validation set 2 is greater than or equal to the preset threshold, the hyperparameter combination is output. If the R of the predicted data 2 is less than the preset threshold, continue to train through the training set within the set number of training times. Among them, the preset value is custom-set between the ranges of the selected evaluation metrics according to project requirements. If the evaluation metric uses the scoring function R 2It can be expressed as 0.4, 0.5, 0.7, etc. If the evaluation index is less than the preset value, return to S44 and adjust the number of iterations and the number of cross-validation folds in the Bayesian optimization algorithm, and continue with hyperparameter optimization until the obtained optimal seepage prediction model meets the expected prediction accuracy requirements.

[0107] Step 6: Use the TPESampler-IKSLS-AdaBoost model obtained in Step 5 for dam seepage prediction.

[0108] Step 6.1: If it is necessary to predict the cumulative seepage volume of the target dam in the next 3 days, taking the first day as the observation day, obtain the corresponding P Q1-10 , L R1-2 , L R2-3 , P H1-3 , P H3-5 , P H5-10 , P H10-15 , C and lnC data through seepage volume and environmental quantity monitoring equipment, and perform normalization processing.

[0109] Step 6.2: According to the observation day date of the cumulative seepage volume to be predicted in the next 3 days of the target dam, determine the key samples in the training set obtained in Step 4, providing a basis for the training of the TPESampler-IKSLS-AdaBoost model.

[0110] Step 6.3: For the TPESampler-IKSLS-AdaBoost model established in Step 5, set the types of hyperparameters to the maximum depth of the tree, the number of base estimators, and the learning rate respectively, and set the value spaces of the hyperparameters to 1 - 7, 50 - 200, and 0.01 - 1 respectively. The number of training times n_trials is 100.

[0111] Step 6.4: For the TPESampler-IKSLS-AdaBoost model established in Step 5, set the types of hyperparameters to the maximum depth of the tree, the number of base estimators, and the learning rate respectively, and set the value spaces of the hyperparameters to 1 - 7, 50 - 200, and 0.01 - 1 respectively. The number of training times n_trials is 100.

[0112] Step 6.5: Input the training set, validation set divided in Step 4, and the key samples determined in Step 6.2 into the model established in Step 5 to determine the key samples in the training set obtained in Step 4, perform model training, and take the optimal hyperparameter combination corresponding to the best training effect in 100 trainings.

[0113] Step 6.6: Input the optimal hyperparameters obtained in Step 6.5 into the IKSLS-AdaBoost model to obtain the optimal IKSLS-AdaBoost model. Input the training set into the optimal IKSLS-AdaBoost model for training to obtain the optimal seepage prediction model for the cumulative seepage flow to be predicted for a certain 3 days in the future of the target dam.

[0114] Step 6.7: Input the data obtained in Step 6.1 into the optimal seepage prediction model for the cumulative seepage flow to be predicted for a certain 3 days in the future of the target dam, and then perform denormalization processing on the output data, which is the cumulative seepage flow of the target dam in the next 3 days.

[0115] In the intelligent dam seepage prediction method introducing the key sample learning strategy of the present invention: The function of Step 2 is to perform noise reduction processing on P Q1-10 , L R1-2 , L R2-3 , P H1-3 , P H3-5 , P H5-10 , P H10-15 and Y in the sample dataset obtained in Step 1 by using the convolutional smoothing filtering method to reduce the influence of abnormal error values on the seepage prediction accuracy.

[0116] The convolutional smoothing filtering method is adopted, and its principle is: By sliding a small window on the data sequence and performing weighted averaging on the data within the window, random noise is reduced and the signal is smoothed.

[0117] The advantage of adopting the convolutional smoothing filtering method is that: When performing noise reduction on the data sequence, this method can effectively remove noise interference while retaining the inherent trend and characteristics of the signal.

[0118] In the novel intelligent dam seepage prediction method introducing the key sample learning strategy of the present invention: The function of Step 5 is to establish the TPESampler-IKSLS-AdaBoost model, introduce the key sample learning strategy specifically for the dam seepage prediction problem, and improve the sample update scheme of the IKSLS-AdaBoost model; In addition, the TPESampler algorithm is used to optimize the hyperparameters of the IKSLS-AdaBoost model.

[0119] The TPESampler-IKSLS-AdaBoost model is adopted, and its principle is: Combining the unique characteristics of the dam seepage prediction problem, the key sample learning strategy is introduced. In addition, the TPESampler optimization algorithm is used to optimize the hyperparameters of the IKSLS-AdaBoost model during the training process, quickly select the optimal hyperparameter combination of the IKSLS-AdaBoost model, and thus establish the optimal IKSLS-AdaBoost model.

[0120] The TPESampler-IKSLS-AdaBoost model is adopted, and its advantages are as follows: In the IKSLS-AdaBoost model, a key sample learning strategy is introduced. Since the external environmental quantities of the dam in similar months within different years have similar effects, resulting in similar seepage effects of the dam, when predicting the cumulative seepage volume to be predicted for a certain period in the future of the target dam, the weights of the samples in the sample dataset whose observation dates of the cumulative seepage volume to be predicted for a certain period in the future of the target dam are close in month are increased for key learning and training, which can greatly improve the accuracy of seepage prediction. By using the good optimization ability of the TPESampler algorithm in nonlinear problems and the high optimization efficiency based on the KD-tree structure, the optimal AdaBoost model hyperparameter combination is quickly selected, thereby improving the prediction ability of the AdaBoost model.

[0121] Example 3: The following is an example to elaborate on the present invention in detail. As Figure 1 、 Figure 2 shown, it is specifically implemented according to the following steps:

[0122] Step 1: Obtain the historical monitoring data Y of the seepage of a certain earth-rock dam and the influencing factor data P Q1-10 、L R1-2 、L R2-3 、P H上2-3 、P H上3-5 、P H上5-10 、P H上10-15 、C, lnC. The historical monitoring data of the seepage of the earth-rock dam is further divided into data sequences at 3-day time intervals, where each data represents the cumulative seepage volume of the earth-rock dam in 3 days, and the corresponding influencing factor seepage observation date is based on the first day of the 3 days. The seepage data sequence of the earth-rock dam and the influencing factors corresponding to each seepage volume data are sorted and combined to construct an initial sample dataset, where the number of samples is 202.

[0123] Step 2: Perform noise reduction processing on P Q1-10 、L R1-2 、L R2-3 、P H1-3 、P H3-5 、P H5-10 、P H10-15 and Y in the sample dataset obtained in Step 1 by using the convolutional smoothing filtering method to obtain a new sample dataset.

[0124] Step 3: Perform standardization processing on various types of data in the new sample dataset obtained in Step 2 to obtain a standardized sample dataset.

[0125] Step 4: Divide the 202 sets of data in the standardized sample dataset constructed in Step 2 into a training set of 141 sets, a validation set of 41 sets, and a test set of 20 sets according to the ratio of 7:2:1.

[0126] Step 5: Establish a new type of intelligent prediction model for dam seepage, TPESampler-IKSLS-AdaBoost, which introduces the key sample learning strategy.

[0127] Step 6: According to the observation date of each sample in the test set, respectively determine the key samples in the training set obtained in Step 4 and make predictions respectively according to Step 6. Take one sample for detailed description. Set three hyperparameters: the maximum depth of the tree, the number of base estimators, and the learning rate. The value spaces of the hyperparameters are set as 1-7, 50-200, and 0.01-1 respectively, and the number of training times n_trials is 100. Input the determined key samples, the training set, and the validation set into the TPESampler-IKSLS-AdaBoost model for training to obtain the optimal hyperparameter combination: the maximum depth of the tree: 6, the number of base estimators: 80, the learning rate: 0.76. Construct the optimal IKSLS-AdaBoost model based on the obtained optimal hyperparameter combination. Input the training set into the optimal IKSLS-AdaBoost model to train the optimal seepage prediction model corresponding to this sample. Input the feature data of this sample into the optimal seepage prediction model corresponding to this sample to obtain the predicted cumulative seepage volume. Compare the predicted values and the actual values of each sample in the test set as Figure 2 shown, and calculate the R 2 between the seepage prediction value and the actual value of the test set is 0.859, and the average training and prediction time is 32.76 seconds, indicating good prediction effect and high efficiency.

Claims

1. An intelligent prediction method for dam seepage incorporating a key sample learning strategy, characterized in that It includes the following steps: S1. Obtain the data obtained by the dam seepage monitoring equipment and the environmental quantity monitoring equipment and conduct preliminary processing, then use the convolutional smoothing filtering method for noise reduction processing to construct a monitoring data set; S2. Based on the monitoring data set, divide the monitoring data set into a training set, a validation set and a test set in a ratio of 7:2:1; S3. Establish an IKSLS-AdaBoost model and conduct iterative training of weak regressors to obtain a new IKSLS-AdaBoost model. Then use the training set and the validation set as inputs for training and learning to obtain the optimal hyperparameter combination of the dam seepage prediction model. Establish a TPESampler-IKSLS-AdaBoost model according to the optimal hyperparameter combination; S4. Use the test set to verify the TPESampler-IKSLS-AdaBoost model and conduct seepage flow prediction.

2. The intelligent prediction method for dam seepage incorporating a key sample learning strategy according to claim 1, characterized in that The S1 includes: S11. Analyze the seepage influence factors for a specific engineering project and obtain historical seepage flow monitoring data through corresponding monitoring equipment; S12. Based on the hysteresis effect, perform difference processing on the historical seepage flow monitoring data by day to obtain a continuous data sequence with a daily time interval; S13. Divide the continuous data sequence at intervals of a custom observation time period to obtain divided data sequences. Each divided data in the divided data sequence represents the historical seepage flow monitoring data within the custom observation time period; S14. Integrate the historical seepage flow monitoring data corresponding to each divided data, and combine it with the divided data sequence to construct an initial sample data set; S15. Use the convolutional smoothing filtering method to input the divided data into the convolutional smoothing filtering model respectively for noise reduction processing to obtain noise-reduced sample data; S16. According to the historical seepage flow monitoring data of the initial sample data set, combine it with the noise-reduced sample data to construct a monitoring data set and perform normalization processing.

3. The intelligent prediction method for dam seepage incorporating a key sample learning strategy according to claim 2, characterized in that, The initial sample data set is expressed as follows: Among them, C i , lnC i is the eigenvalue, representing the data of the initial target object seepage influence factors; Y i is the target variable, representing the preferred sample data; N represents the number of the initial target object seepage influence factors; M represents the number of data groups in the sample dataset.

4. According to the dam seepage intelligent prediction method introducing a key sample learning strategy described in claim 2, characterized in that: The specific expression of the normalization processing is as follows: X, = (X - X min ) / (X max - X min ) ; Among them, X is the value after normalization, X is the original data, X max is the maximum value in the dataset, X min is the minimum value in the dataset.

5. The intelligent prediction method for dam seepage incorporating the key sample learning strategy according to claim 1, characterized in that, The specific expressions of the training set, the validation set and the test set in S2 are: Validation set: Test set:

6. The intelligent prediction method for dam seepage incorporating a key sample learning strategy according to claim 1, wherein, The S3 includes the following steps: S31. Initialize the weights based on the training set. Take the sample distribution of the training set in the initial state as Dt1, the number of samples in the dataset as m, and initialize the weight of each sample X i to 1 / m. Dt1 is used for the training of the first weak regressor h1, and Dt t is used for the training of the first weak regressor ht, and so on; S32. Perform T iterations in a loop. In each iteration, the weak learner number is t, t ∈ {1, 2, 3,..., T}. This step is a loop body; S321. On the basis of the sample weight distribution being Dt t (X), train the weak classifier h on the training set D t ; Calculate the maximum error E of the regressor h t on the training set D t ; According to E t calculate the squared error e of h t for each sample, so as to obtain the error rate e of the current weak regressor h ti and finally calculate the weight w of the weak regressor h t ; t t t ;​​ S322. Update the sample weight distribution of the training set. The calculation formula for the maximum weight of the training set samples in the next iteration of training is expressed as follows: Among them, Dt t+1 (X)max represents the maximum weight of the dataset samples for the next iterative training; S323. Update all the weights of the key samples during the next iteration training to Dt t+1 (X)max. That is, the weight calculation formula for other samples is as follows: Among them, Z t is a normalization factor, and its calculation formula is: If X i is a key sample, then is Dt t+1 (X)max, otherwise is S324. Let t = t + 1, and return to S321 to continue the iteration. After T rounds of iteration, finally obtain the IKSLS-AdaBoost model; S33. Establish a function. Each time this function is called, the AdaBoost models corresponding to different hyperparameter combinations will be trained once and the training effect R will be returned. 2 : S331. Establish a function. First, set the types and search spaces of the hyperparameters of the IKSLS-AdaBoost model in the function; S332. According to the value-taking rule of the TPESampler optimization algorithm, each time the function is called, a set of hyperparameter values of the IKSLS-AdaBoost model is taken, and a new IKSLS-AdaBoost model is established; S333. Use the training set as input, train based on the new IKSLS-AdaBoost model, then use the validation set to verify the training effect, calculate and return the R between the predicted data and the actual data of the validation set 2 ; S34. Call the established function, set the number of training times n_trials, and set the optimal hyperparameters to be the optimal hyperparameter combination in n_trials times of training, to complete the establishment of the TPESampler-KSLS-AdaBoost model.

7. A dam seepage intelligent prediction method introducing a key sample learning strategy according to claim 6, characterized in that: The obtaining of the hyperparameter combination corresponding to the optimal training effect in n_trials training includes calculating the R of the predicted data of the validation set 2 and comparing it with a preset threshold. If the R of the validation set 2 is greater than or equal to the preset threshold, the hyperparameter combination is output. If the R of the predicted data 2 is less than the preset threshold, continue to train through the training set within the set number of training times.

8. A dam seepage intelligent prediction method introducing a key sample learning strategy according to claim 6, characterized in that: The types of hyperparameters include the number of decision trees, the learning rate, the maximum depth of the tree, the minimum number of samples required to split a node, and the minimum number of samples that a leaf node should have.

9. The intelligent prediction method for dam seepage incorporating a key sample learning strategy according to claim 7, wherein, The S4 includes: Input the test set into the optimal seepage prediction model for verification. When predicting the cumulative seepage volume of the target dam in the future time period, taking the first day as the observation day, through the dam seepage monitoring equipment and environmental quantity monitoring equipment, obtain the historical seepage volume monitoring data corresponding to the seepage volume on the observation day, perform normalization processing, and then perform seepage volume prediction.

Images