Prediction method of leaching behavior of high-energy glass solidification based on temporal convolutional network

Abstract

The present invention provides a method for predicting the leaching behavior of high-radiation glass solids based on a time-series convolutional network, comprising the following steps: (1) data preprocessing; (2) prediction model design; and (3) element leaching behavior prediction. The present invention utilizes a time-series convolutional network as the prediction model structure, effectively improving the accuracy of predicting the leaching behavior of high-radiation glass solids. The model is trained using the ALTGLASS international database as a training set, avoiding the problems of inconsistent experimental conditions and insufficient data volume that have arisen in previous studies when obtaining training set data from literature or experiments.

Description

Technical Field

[0001] The present invention provides a method for predicting the leaching behavior of a high-radiation glass solid body based on a temporal convolutional network, and belongs to the technical field of nuclear safety assessment. Background Art

[0002] High-level radioactive glass solids come into contact with groundwater during long-term deep geological disposal, leading to the leaching of radionuclides. Therefore, studying and predicting the leaching behavior of glass solids is an important part of nuclear safety assessment. Currently, the commonly used method for studying and predicting leaching behavior is the traditional experimental method of preparing samples and conducting leaching experiments of varying durations. This method has the disadvantages of a long experimental cycle and the inability to generalize leaching results to other components. Existing models describing the leaching of glass solids, such as the solution analysis model (JNUCL MATER, 2001, 298(1):112-124) and the geochemical model (Waste Management, 2001, 21(8):741-752), are unable to fully reflect the complexity of the leaching behavior of glass solids. Summary of the Invention

[0003] In response to the above technical problems, the present invention uses a time-series convolutional network model based on a machine learning method to train a prediction model for the leaching behavior of high-level radioactive glass solids, providing new technical means and ideas for the research and prediction of the leaching behavior of high-level radioactive glass solids.

[0004] The present invention uses the international large-scale high-level nuclear waste glass leaching database ALTGLASS to train and test the time series convolutional network model, and obtains a leaching behavior prediction model that can make relatively accurate predictions on the leaching behavior of 11 elements in high-level glass solids. Traditional experimental methods and descriptive models are difficult to show the complexity of the leaching behavior of glass solids, while the data-driven model based on machine learning is a semi-empirical model developed based on the existing database, which can model complex situations. Among them, models such as artificial neural networks and support vector machines have achieved success in the study of the dissolution behavior of phosphate glass and aluminosilicate glass. However, methods such as artificial neural networks and support vector machines have problems such as insufficient processing capabilities for large-scale data, gradient vanishing and overfitting, and therefore often have limited prediction accuracy. As a deep learning model, the time series convolutional network has strong data feature extraction and fitting capabilities, and can provide better prediction capabilities for the leaching behavior of high-level glass solids. The ALTGLASS database also contains 2,600 static glass dissolution data sets for approximately 450 different glass compositions, including high-level and low-level waste vitrification. All experiments were conducted using product consistency testing (PCT, ASTM C1285A and B) methods, resulting in high reliability. Therefore, using the ALTGLASS database to train a temporal convolutional network is expected to yield a model for predicting the leaching behavior of high-level vitrification that can accurately predict the leaching behavior of high-level vitrification.

[0005] The specific technical solution of the present invention is:

[0006] The method for predicting the leaching behavior of high-energy glass solidification bodies based on a temporal convolutional network includes the following steps:

[0007] (1) Data preprocessing

[0008] The static leaching data of high-energy glass solids obtained in deionized water at 90℃±2℃ were selected from the ALTGLASS dataset. The element normalized mass loss Q was calculated according to the concentration of the element in the leachate using formula (1): i :

[0009]

[0010] Where: Q i The unit is g / m 2 ; C i is the concentration of the element in the leachate, in g / L; f i is the mass fraction of the element in the glass sample; SA / V is the ratio of the sample surface area to the leachate volume, in m 2 / L. In order to process input quantities of different dimensions, the data is normalized using formula (2):

[0011]

[0012] Where: Y i is the normalized result; x i is the current data; x min and x max Before training, 64%, 16%, and 20% of the data are randomly divided as training, validation, and test sets, respectively.

[0013] (2) Prediction model design

[0014] The input features include leaching duration, SA / V, sample density, leachate pH, and sample composition, totaling 43 dimensions. The output features are the predicted leaching values for 11 elements. The prediction model, written in Python, consists of an input layer, a TCN layer (using the Keras package), and a Dense layer. The TCN layer uses 43 time steps. The temporal convolutional network model is trained using the training set as input. The validation set's MAE (meaningful error estimation) is used to assess training performance and avoid overfitting. After training, a prediction model for the leaching behavior of the element is obtained.

[0015] In order to evaluate the prediction accuracy of the model, the trained model was tested with the test set data, and the mean absolute error (MAE) and the coefficient of determination (R) were calculated using the model calculation value and the test set data through formula (3) and formula (4). 2 ) to quantitatively analyze the test results.

[0016]

[0017]

[0018] (3) Element leaching behavior prediction

[0019] Use Python to read the trained element leaching behavior prediction model, normalize the input features to be predicted as above, and then submit them to the model for running to obtain the element leaching prediction value.

[0020] The present invention adopts a temporal convolutional network as the prediction model structure. It is an improved model based on 1D convolution for timing problems. In addition, it also adopts residual convolution skip layer connection, 1×1 convolution and other designs, which can effectively improve the accuracy of predicting the leaching behavior of high-level glass solidification bodies.

[0021] The ALTGLASS model was trained using the training data set. ALTGLASS International researchers provided an experimental data archive. All data was averaged from three experiments conducted using the short-term and long-term product consistency testing (PCT, ASTM C1285 A and B) methods. The database is extensive, containing 2,600 static glass dissolution data points for approximately 450 different glass compositions. This database meets the requirements for model training and avoids the inconsistent experimental conditions and insufficient data required to obtain training data from literature or experiments in previous studies. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is the normalized test set determination coefficient and mean absolute error obtained by the present invention using the temporal convolutional network prediction;

[0023] Figure 2 3. It is a scatter plot comparison of the calculated values and experimental values of the time series convolution prediction model training set and test set for B, Si and Al elements in the embodiment; DETAILED DESCRIPTION

[0024] The specific technical solutions of the present invention are described with reference to the embodiments.

[0025] The trained model was used to predict the leaching behavior of Si, B, Ca, Cr, Li, Na, K, Al, Fe, Mo and Mg elements using the test set, and the mean absolute error (MAE) and determination coefficient (R) were used to compare the calculated values with the experimental values of the test set. 2 ) to evaluate the accuracy of the model. The closer MAE is to 0, the better the model prediction accuracy; R 2 The closer it is to 1, the better the model prediction accuracy. 2 The closer the value is to 1 and the lower the MAE value is, the better the model prediction performance is considered.

[0026] MAE and R on the normalized test set 2 As a result, the results are as follows Figure 1 As shown. Figure 1 From (a) and (b) in the figure, we can see that for Al element, the R 2 The value is about 0.95, and the MAE error value is about 0.0120. 2 The values are 0.91, 0.94 and 0.95 respectively, and the MAE values are 0.0272, 0.0160 and 0.0251 respectively. 2 The R values of K, Mo, and Na are 0.42 and 0.0182 respectively. 2The values are 0.67, 0.86 and 0.63 respectively, and the MAE values of the three elements are 0.0071, 0.0256 and 0.0658 respectively. 2 is 0.96, and MAE is 0.0022. 2 and MAE were 0.786 and 0.0212 respectively. 2 The R values of the model for Al, B, Cr, Li, and Si are 0.75 and 0.0056 respectively. 2 The values are all above 0.9, indicating that the model can provide high prediction accuracy for a variety of elements.

[0027] Table 1 shows the normalized test set determination coefficient and mean absolute error of Al, Li, B, and Ca elements predicted by artificial neural network, time series convolutional network, and convolutional neural network. From the results, it can be found that the R 2 Both are higher than those of artificial neural networks and convolutional neural networks, and the MAE is lower than those of artificial neural networks and convolutional neural networks, indicating that the prediction accuracy of the time series convolutional network is better than that of the artificial neural networks and convolutional neural networks used in existing studies.

[0028] Table 1 Normalized test set determination coefficients and mean absolute errors of Al, Li, B, and Ca elements predicted using artificial neural networks, temporal convolutional networks, and convolutional neural networks

[0029]

[0030] In this embodiment, the element leaching value data and input features are input into the model as a training set for training, so as to obtain a prediction model for the leaching behavior of the element. Figure 2 (a) Figure 2 (b) Taking B, Si, and Al as examples, a scatter plot of the calculated and experimental values of the training and test sets of the time series convolution prediction model is shown. The figure represents B, Si, and Al elements with spheres, triangles, and diamonds, respectively. The dotted line represents the ideal fitting line where the experimental and calculated values are completely consistent. The closer the data point is to the line, the better the agreement between the calculated and experimental values at that point. The more points in the data set are close to the line, the higher the prediction accuracy of the data set. It can be seen from the figure that the training set data points of B, Si, and Al are in good agreement with the ideal fitting line, indicating that the time series convolution model did not suffer from underfitting or overfitting during the training process of the leaching behavior of B, Si, and Al elements. The distribution of the test set data points is close to the ideal fitting line, indicating that the calculated and experimental values of the test set are in good agreement. The prediction model shows excellent prediction accuracy for the leaching behavior of B, Si, and Al.

Claims

1. A method for predicting the leaching behavior of high-energy glass solidification bodies based on a temporal convolutional network, characterized in that: The following steps are involved: (1) Data preprocessing The static leaching data of high-energy glass solids obtained in deionized water at 90℃±2℃ were selected from the ALTGLASS dataset; the element normalized mass loss Q was calculated according to the concentration of the element in the leachate using formula (1): i : Where: Q i The unit is g / m 2 ; C i is the concentration of the element in the leachate, in g / L; f i is the mass fraction of the element in the glass sample; SA / V is the ratio of the sample surface area to the leachate volume, in m 2 / L; the data are normalized using formula (2): Where: Y i is the normalized result; x i is the current data; x min and x max are the minimum and maximum values of the data respectively; Before training begins, the data is randomly divided into training set, validation set and test set respectively; (2) Prediction model design The 43-dimensional data including leaching duration, SA / V, sample density, leachate pH, and sample components were used as input features, and the element's leaching prediction value was used as the output feature. The prediction model was written in Python and consisted of an input layer, a Keras package tcn layer, and a dense layer. The tcn layer used 43 time steps. The training set was used as the input data set to train the temporal convolutional network model. During the training process, the validation set MAE value was used to judge the training effect to avoid overfitting. After training, a prediction model for the element's leaching behavior was obtained. (3) Element leaching behavior prediction Use Python to read the trained element leaching behavior prediction model, perform the normalization processing on the input feature data set required for leaching behavior prediction as above, and then submit it to the model to obtain the predicted value of element leaching.

2. The method for predicting leaching behavior of high-energy glass solidified bodies based on a temporal convolutional network according to claim 1, characterized in that: In step (1), the preprocessed normalized element leaching dataset is randomly divided into 64%, 16% and 20% of the data as the training set, validation set and test set, respectively.

3. The method for predicting leaching behavior of high-energy glass solidified bodies based on a temporal convolutional network according to claim 1, characterized in that: Step (2) also includes evaluating the prediction accuracy of the predicted model: the trained model is tested with the test set data, and the mean absolute error (MAE) and the coefficient of determination (R) are calculated using the model calculation value and the test set data through formulas (3) and (4). 2 To quantitatively analyze the test results: In the formula, n represents the number of samples, y i represents the experimental value of the i-th sample, represents the predicted value of the i-th sample, Represents the average of the sample experimental values.

Images