Mass concrete temperature prediction method and terminal equipment

Abstract

The present invention belongs to the technical field of temperature prediction, and discloses a method for predicting the temperature of large-volume concrete and a terminal device. The method includes obtaining the temperature prediction value of each measuring point of the large-volume concrete predicted by the temperature mechanism model, and determining the temperature error between the temperature prediction value and the corresponding temperature actual measurement value; training a machine learning model according to the temperature influencing factor data and the temperature error of the large-volume concrete; inputting the measured temperature influencing factor data into the trained machine learning model to obtain the error compensation prediction value; and correcting the temperature prediction value using the error compensation prediction value. The present invention is based on the prediction result of the temperature mechanism model, and uses the error compensation prediction value determined by the machine learning model considering uncertain factors to correct the prediction result of the temperature mechanism model, so that the final temperature prediction value is more accurate, and provides reliable data support for the arrangement of the construction schedule and the formulation of temperature crack prevention measures.

Description

Technical Field

[0001] The invention belongs to the technical field of temperature prediction, and specifically discloses a mass concrete temperature prediction method and terminal equipment. Background Art

[0002] Mass concrete has become one of the widely used materials in construction, but its temperature changes during the construction period may lead to the formation of temperature cracks, affecting the safety of the structure. Therefore, it is necessary to accurately predict the temperature change process of mass concrete and control it.

[0003] Hydration heat is one of the important factors affecting temperature change. Existing technologies use the adiabatic temperature rise equation as a temperature mechanism model to simulate and predict the temperature change of large-volume concrete. Although this mechanism model analysis can describe the temperature rise trend during the concrete solidification process, the heat conduction equation, boundary conditions and external environment of the mechanism model are simplified to a certain extent during actual calculations, resulting in inaccurate predictions in local areas of the prediction results, which cannot meet the needs of dynamic and efficient analysis during the construction process. Summary of the invention

[0004] The purpose of the present invention is to provide a method and terminal device for predicting the temperature of large-volume concrete, so as to solve the technical problem that the existing temperature prediction method requires a certain degree of simplification of the heat conduction equations, boundary conditions and external environment of the mechanism model, resulting in inaccurate prediction of local areas in the prediction results.

[0005] The first aspect of the present invention discloses a method for predicting the temperature of a large volume of concrete, comprising:

[0006] Step 1: Obtain the temperature prediction value of each measuring point of the mass concrete predicted by the temperature mechanism model, and determine the temperature error between the temperature prediction value and the corresponding temperature measured value.

[0007] Step 2: training a machine learning model based on the temperature influencing factor data of mass concrete and the temperature error.

[0008] Step 3: Input the measured temperature influencing factor data into the trained machine learning model to obtain the error compensation prediction value.

[0009] Step 4: Correct the temperature prediction value using the error compensation prediction value.

[0010] Preferably, the machine learning model is an XGBoost model; then the step 2 specifically comprises: taking the temperature influencing factor data of the mass concrete as input and the temperature error as output to train the XGBoost model.

[0011] Preferably, the XGBoost model includes multiple hyperparameters; the multiple hyperparameters are optimized using the longicorn beard search algorithm; the fitness function in the longicorn beard search algorithm is determined based on the correlation coefficient and mean square error between the temperature prediction value and the corresponding temperature measured value.

[0012] Preferably, the step size adjustment method of the longhorn beard search algorithm in the iterative process is:

[0013] Step S1, obtaining the standard deviation of the fitness value within a preset number of iterations n during the process of adjusting the step size using the step size adjustment formula.

[0014] Step S2: when the standard deviation is less than or equal to a preset threshold, the number of iterations is increased by 1; when n≥the maximum number of iterations m, the step length is expanded to a preset multiple of the step length corresponding to the current number of iterations and the number of iterations n is reset to zero; when the standard deviation is greater than the preset threshold or n<m, the step length is adjusted using the step length adjustment formula; steps S1 and S2 are repeated until the maximum number of iterations is met or the preset accuracy requirement is met.

[0015] Preferably, before step 1, the method further includes: step A, performing parameter inversion on thermal parameters in the original temperature mechanism model to obtain the temperature mechanism model.

[0016] Preferably, the step A specifically comprises:

[0017] Step A1: determining a mapping relationship between the temperature of the mass concrete and the thermal parameters.

[0018] Step A2: determining the inversion value of the thermal parameter according to the measured temperature value and the mapping relationship, and substituting the inversion value into the original temperature mechanism model to obtain the temperature mechanism model.

[0019] Preferably, the step A1 specifically includes:

[0020] Obtain multiple sets of design parameter values ​​and corresponding design temperature values ​​of the thermal parameters to be inverted in the original temperature mechanism model.

[0021] Construct an inversion proxy model consisting of a neural network model and a regression prediction model.

[0022] The design temperature value is used as input and the design parameter value is used as output to train the inversion proxy model to obtain a mapping relationship between the temperature of the mass concrete and the thermal parameter.

[0023] Preferably, the inversion proxy model is a Stacking ensemble model; the base learners in the Stacking ensemble model are a neural network model and a regression prediction model; and the meta learner in the Stacking ensemble model is a linear regression model.

[0024] Preferably, the neural network model is a BP neural network model; the activation function of the BP neural network model includes a steepness factor, and the steepness factor is ≥1.

[0025] The second aspect of the present invention discloses a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned mass concrete temperature prediction method when executing the computer program.

[0026] Compared with the prior art, the mass concrete temperature prediction method and terminal device of the present invention have the following beneficial effects:

[0027] The present invention proposes a parallel fusion model, which is based on the prediction results of the temperature mechanism model and uses a machine learning model that takes uncertain factors into consideration to correct local errors, so that the obtained temperature prediction value is more accurate.

[0028] In order to improve the accuracy of the temperature prediction value of the temperature mechanism model, the present invention inverts the thermal parameters therein. However, during the construction of large-volume concrete, uncertain factors such as environmental conditions and meteorological information will cause the temperature mechanism model to have deviations even after parameter inversion optimization. The single data model lacks an in-depth understanding of the hydration thermal reaction of large-volume concrete, and the limited amount of data makes it impossible to build a high-accuracy temperature prediction model. Therefore, after obtaining the temperature prediction value using the temperature mechanism model after parameter inversion, the present invention corrects the temperature prediction value through the error compensation prediction value determined by the machine learning model to obtain a more accurate temperature prediction result, thereby providing reliable data support for the arrangement of construction progress and the formulation of temperature crack prevention measures.

[0029] The mass concrete temperature prediction method of the present invention has the advantages of solving problems such as small samples, nonlinearity, and high dimensions, and reduces calculation time and resource consumption while ensuring accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 The figure is a flow chart of a method for predicting temperature of mass concrete according to an embodiment of the present invention.

[0031] Figure 2 The present invention is a schematic diagram of a step size adjustment method in an improved longicorn beetle whisker search algorithm during an iteration process in an embodiment of the present invention.

[0032] Figure 3 It is the iterative optimization curve of BAS and IBAS in the embodiment of the present invention.

[0033] Figure 4 This is a graph of the prediction results of the IBAS-XGBoost model in an embodiment of the present invention.

[0034] Figure 5 This is a diagram of the prediction results of the BAS-XGBoost model in an embodiment of the present invention.

[0035] Figure 6 This is a graph of XGBoost model prediction results in an embodiment of the present invention.

[0036] Figure 7 This is a diagram of the prediction results of the SVR model in an embodiment of the present invention.

[0037] Figure 8 This is a diagram of the prediction results of the BP neural network model in an embodiment of the present invention.

[0038] Fig. 9 This is the temperature prediction result of point A on the water-facing surface by the fusion model of the temperature mechanism model and the machine learning model in an embodiment of the present invention.

[0039] Fig.10 This is the temperature prediction result of point B on the water-facing surface by the fusion model of the temperature mechanism model and the machine learning model in an embodiment of the present invention.

[0040] Fig.11 This is the temperature prediction result of point A on the back water surface by the fusion model of the temperature mechanism model and the machine learning model in the embodiment of the present invention.

[0041] Fig.12 This is the temperature prediction result of point B on the back water surface by the fusion model of the temperature mechanism model and the machine learning model in the embodiment of the present invention.

[0042] Fig.13 This is the temperature prediction result of the fusion model of the temperature mechanism model and the machine learning model for the back water point C in the embodiment of the present invention.

[0043] Fig.14 This is the temperature prediction result of the fusion model of the temperature mechanism model and the machine learning model for the back water point D in the embodiment of the present invention.

[0044] Fig.15 This is a comparison diagram of the monitoring value and the inversion value output of point A on the water-facing surface in an embodiment of the present invention.

[0045] Fig.16 This is a comparison chart of predicted values ​​and designed values ​​based on different models of total adiabatic temperature rise of cement in an embodiment of the present invention.

[0046] Fig.17 The hydration heat coefficient in the embodiment of the present invention is Comparison chart of predicted values ​​and designed values ​​of different models.

[0047] Fig.18 The hydration heat coefficient in the embodiment of the present invention is Comparison chart of predicted values ​​and designed values ​​of different models. DETAILED DESCRIPTION

[0048] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention. However, it should be clear to those skilled in the art that the present invention may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present invention.

[0049] A first aspect of an embodiment of the present invention discloses a method for predicting temperature of a large volume of concrete. Figure 1 As shown, including:

[0050] Step 1: Obtain the temperature prediction value of each measuring point of the mass concrete predicted by the temperature mechanism model, and determine the temperature error between the temperature prediction value and the corresponding temperature measured value.

[0051] The temperature mechanism model of the embodiment of the present invention is specifically an adiabatic temperature rise equation. Since the calculation result of the double exponential adiabatic temperature rise equation is highly consistent with the actual monitoring result, the embodiment of the present invention uses the double exponential adiabatic temperature rise equation as the temperature mechanism model, as shown in formula (1):

[0052] (1)

[0053] In the formula, for The temperature of the mass concrete at the moment; is the total adiabatic temperature rise of cement; and Both are hydration heat coefficients, which are related to the type of cement.

[0054] In the study of temperature prediction of large-volume concrete, the mechanism model often has certain errors due to necessary simplification. This error not only comes from the simplification of the model itself, but is also affected by subjective and objective uncertainty factors, which make it difficult for the mechanism model to meet the needs of dynamic and efficient analysis during the construction process. In order to solve this problem, the embodiment of the present invention constructs a machine learning model. The machine learning model can reveal the complex correlation between the temperature change of large-volume concrete and its influencing factors through massive data, has the ability to handle highly nonlinear problems, and has strong generalization ability, which can achieve efficient analysis and online deployment, thereby providing more accurate prediction results.

[0055] Step 2: Train the machine learning model based on the temperature influencing factor data and temperature error of mass concrete. Specifically, use the temperature influencing factor data of mass concrete as input and the temperature error as output to train the machine learning model.

[0056] There are many factors that affect the temperature change process of large-volume concrete during construction, such as cement mix ratio, location information of monitoring points, concrete density, specific heat, thermal conductivity, concrete heat dissipation coefficient to air (related to wind speed), ambient temperature and humidity, and also information such as the flow rate of cooling water when water is passed through the large-volume concrete for cooling.

[0057] The embodiment of the present invention can use the data corresponding to the above-mentioned temperature influencing factors to train the machine learning model. However, if the above-mentioned temperature influencing factors are directly used as input variables of the machine learning model without analysis, the structure of the machine learning model will be complicated and even affect its prediction performance. Therefore, in order to improve the accuracy of the prediction results, it is necessary to select key features from these complex influencing factors as input variables of the machine learning model, thereby simplifying the structure of the machine learning model.

[0058] The Maximum Information Coefficient (MIC) is based on the mutual information theory and is a combination of information theory and probability statistics. It is used to detect non-functional relationships between different variables and quantify the strength of the dependencies between these variables. The advantages of MIC are its universality and balance: universality enables MIC to reveal linear and nonlinear relationships between variables; and balance enables MIC to compare the strength of relationships between different variables from a vertical or horizontal perspective in different research scenarios.

[0059] The embodiment of the present invention uses MIC to measure the correlation between the temperature change of mass concrete and various temperature influencing factors. Assume that there is a feature set of n samples , where c represents the specific category, namely temperature, and the number of features is m. Any feature The correlation with category c is defined by the MIC value, which varies between 0 and 1. The larger the value, the stronger the correlation. According to the MIC value, the features with strong correlation are retained, and the features with weak correlation or irrelevant are deleted, so that the temperature influencing factors with strong correlation with the temperature change of mass concrete can be obtained, and then the data of the determined temperature influencing factors are obtained and input into the machine learning model.

[0060] The above-mentioned machine learning model can be a BP neural network model or an XGBoost model; since the XGBoost model has excellent learning ability and flexibility, can automatically process missing samples, and has unique advantages in scenarios such as limited training samples and short training time, the embodiment of the present invention preferably uses the XGBoost model.

[0061] XGBoost is a distributed gradient boost decision tree algorithm that uses decision trees as the basic model. The basic principle is to build a strong learner for a given data set by iteratively training k classification and regression trees (CART). Each leaf node of the tree has a prediction score, and each sample is assigned to a leaf node according to its characteristics. Its prediction score is the sum of the leaf node scores of all trees.

[0062] XGBoost has many hyperparameters such as n_estimators, max_depth and learning_rate. The settings of these parameters directly affect the efficiency and accuracy of the XGBoost model. Hyperparameter optimization involves finding the optimal solution in a high-dimensional space. If there are too many optimized parameters, the model may become more complex, affecting the model's computational efficiency and thus reducing the overall performance. At present, most studies use artificial experience, grid search methods or Bayesian search methods to explore the best hyperparameters for XGBoost. However, the use of artificial experience to optimize hyperparameters requires a large number of experiments and rich empirical knowledge. The grid search method is easily constrained by dimensionality, has a limited search range, and is difficult to find the optimal parameters. Although the Bayesian optimization algorithm can effectively reduce the time for hyperparameter optimization, it may fall into the local optimum and it is difficult to find the global optimal hyperparameter combination.

[0063] In order to overcome the problems existing in the above-mentioned hyperparameter optimization method, the embodiment of the present invention adopts the beetle whisker search algorithm to optimize multiple hyperparameters.

[0064] Beetle Antennae Search (BAS) is a bionic heuristic intelligent optimization algorithm. The algorithm simulates the process of beetles searching for food, using the left and right antennae of an individual beetle to sense the strength of the food taste signal, and adjusts the direction of advance according to the signal difference until it reaches the position with the strongest signal. The algorithm does not require the specific form and gradient information of the objective function, and is suitable for solving complex optimization problems. The algorithm has a simple structure and small amount of calculation. During the foraging process, the beetle moves by searching for the taste signal emitted by the food, and uses the left and right antennae for detection and perception. The steps of the algorithm are as follows:

[0065] (1) Set the initial parameters of the algorithm, including the relationship factor between the antenna distance and the step length, the initial step length, the number of iterations, and the error accuracy. The value of the initial step length is generally equivalent to the search interval range, which represents the distance the antenna moves in each iteration.

[0066] (2) Define the initial center of mass position of the longhorn beetle, which is a set of randomly generated multidimensional vectors whose dimensions are determined by the dimension of the problem to be solved.

[0067] (3) Calculate the positions of the left and right antennae, and the orientation of the beetle's head is a randomly created direction vector.

[0068] (4) Select the fitness function, calculate the fitness values ​​at the positions of the left and right antennae, and move the center of mass of the longhorn beetle.

[0069] (5) Considering the local search capability of the algorithm, the antenna distance and step size need to be updated.

[0070] (6) Determine whether the accuracy or the maximum number of iterations has been achieved. If not, return to (3) to continue running. Otherwise, the centroid position of the longhorn beetle at this time is the global optimal solution.

[0071] In the above-mentioned standard BAS algorithm, the optimization process only involves a single individual longhorn beetle, which causes the algorithm to converge quickly. However, this method has limitations when dealing with complex function optimization problems. Since the longhorn beetle randomly selects the direction of movement in each iteration, it cannot be guaranteed that the fitness value will be improved after each update, and the search range is limited, which may lead to reduced algorithm performance and difficulty in locking the global optimal solution. In order to overcome these limitations, the embodiment of the present invention introduces a longhorn beetle group optimization strategy, that is, multiple longhorn beetles explore simultaneously in different directions, which expands the search space, enhances the algorithm's optimization ability in complex problem space, and helps to avoid falling into the dilemma of local optimal solutions. The longhorn beetle population can be represented by a matrix:

[0072] (2)

[0073] (3)

[0074] Where: Indicates the problem dimension corresponding to each longhorn beetle ; For a single longhorn beetle, , is the population size of the longhorn beetle. The corresponding longhorn beetle fitness value can be expressed as:

[0075] (4)

[0076] (5)

[0077] Where: Each value in represents the fitness value corresponding to each longhorn beetle; Each value in represents the problem dimension in each longhorn beetle The fitness value of .

[0078] Furthermore, the embodiment of the present invention also improves the fitness function in the longicorn beard search algorithm, and the fitness function is determined according to the correlation coefficient and the mean square error between the temperature prediction value and the corresponding temperature measured value.

[0079] Exemplarily, the fitness function is determined according to the following formula (6):

[0080] (6)

[0081] In the formula, For fitness, is the correlation coefficient between the predicted temperature value and the corresponding measured temperature value, is the mean square error between the predicted temperature value and the corresponding measured temperature value.

[0082] From the principle of BAS algorithm, it may fall into the local optimal solution, so some strategies need to be adopted to improve the optimization effect. At present, the most commonly used one is variable step size, that is, dynamically adjusting the step size of the algorithm according to the number of iterations. In the initial stage, the step size is set to be equivalent to the width of the solution space. As the beetle moves, the step size gradually decreases. This can avoid the possibility of the beetle falling into the local optimal solution to a certain extent, ensure that the beetle can walk through the entire solution search space, and also speed up the convergence of the algorithm, so that the algorithm can obtain better results within a limited time. The variable step size method is that as the number of iterations increases, the beetle's moving step size shows an exponential decay trend. In the early stage of the iteration, the beetle's step size is large, which helps to perform global search and avoid falling into the local optimal solution too early; while in the later stage of the iteration, the step size is gradually shortened, which is conducive to detailed local search, thereby enhancing the accuracy of optimization. However, relying solely on the variable step size strategy is not enough to ensure the excellent convergence characteristics of the beetle whisker search algorithm. In the later stages of the iteration process, the beetle's step length and the distance between its two antennae will gradually decrease, which limits its full exploration of the search space and may cause the algorithm to fall into a local optimal solution. In order to improve the algorithm's later global search capability and avoid premature convergence, it is necessary to further optimize the algorithm structure. The embodiment of the present invention proposes the following method for adjusting the step length of the beetle's antennae search algorithm during the iteration process: Figure 2 As shown, the algorithm still has a relatively wide search space in the later stage, specifically:

[0083] First, set the maximum number of stable iterations (m), the threshold for sufficiently stable fitness (threshold), and the number of consecutive stable iterations (n). Use an array to store the fitness value of each iteration, and then proceed as follows:

[0084] Step S1, obtaining the standard deviation (Std) of the fitness value within a preset number of iterations n during the process of adjusting the step size using the step size adjustment formula.

[0085] The step size adjustment formula used in the embodiment of the present invention is as follows:

[0086] (7)

[0087] In the formula, For the The step size of the iteration; and are constants, and in the embodiment of the present invention, they are taken as 1.1 and 0.4 respectively; is the number of iterations.

[0088] Step S2, when the standard deviation is less than or equal to the preset threshold, the number of iterations is increased by 1, and the number of iterations is compared with the maximum stable number of iterations. When n≥m, it means that the fitness value may not be updated and the algorithm is trapped in the local optimal solution. Therefore, the step length is expanded to a preset multiple of the step length corresponding to the current number of iterations as the step length to jump out of the local optimal solution. By appropriately increasing the step length of the longhorn beetle, the local search ability of the longhorn beetle is improved and its optimization space is expanded; then after the step length is expanded, the number of iterations n is reset to zero, and the step length of the longhorn beetle is changed again according to formula (7); when the standard deviation is greater than the preset threshold or n<m, the step length is adjusted using formula (7); steps S1 and S2 are repeated until the maximum number of iterations is met or the preset accuracy requirement is met. The above preset multiple can be 2 times.

[0089] Step 3: Input the measured temperature influencing factor data into the trained machine learning model to obtain the error compensation prediction value.

[0090] Step 4: Use the error compensation prediction value to correct the temperature prediction value.

[0091] For example, the final temperature prediction value is as follows:

[0092] (8)

[0093] In the formula, is the temperature prediction value obtained in step 1, is the error compensation prediction value obtained in step 3.

[0094] The present invention utilizes a machine learning model to consider uncertain factors not involved in the mechanism model, such as actual air temperature, inlet and outlet water temperatures, water flow rate, etc., and eliminates uncertainties by performing error compensation and correction on the mechanism model, thereby achieving the goal of accurately evaluating the temperature changes of large-volume concrete.

[0095] In order to accurately predict the temperature changes of large-volume concrete during the construction period, it is necessary to clarify the changing law of hydration heat. Therefore, the selection of concrete adiabatic temperature rise parameters in the temperature mechanism model is particularly critical. However, due to the influence of test conditions, construction technology, construction environment and other factors, the adiabatic temperature rise parameters selected in the design stage experiment have certain errors and fail to fully, accurately and truly reflect the cement hydration heat process in the construction stage, resulting in deviations in the numerical simulation results of the temperature field and may even deviate from the actual situation. Therefore, the embodiment of the present invention uses a parameter inversion method to determine the concrete adiabatic temperature rise parameters.

[0096] The embodiment of the present invention further includes, before step 1:

[0097] Step A: performing parameter inversion on the thermal parameters in the original temperature mechanism model to obtain the temperature mechanism model, specifically including:

[0098] Step A1: determining the mapping relationship between the temperature and thermal parameters of the mass concrete, specifically including:

[0099] Step a: obtaining multiple sets of design parameter values ​​and corresponding design temperature values ​​of thermal parameters to be inverted in the original temperature mechanism model.

[0100] The original temperature mechanism model in the embodiment of the present invention is run in finite element software, wherein the original temperature mechanism model is shown in the above formula (1).

[0101] The thermal parameter to be inverted determined in the embodiment of the present invention is the total adiabatic temperature rise of cement in formula (1): and hydration heat coefficient , .

[0102] The multiple sets of design parameter values ​​for the thermal parameters to be inverted in the above-mentioned original temperature mechanism model are specifically:

[0103] Get the total adiabatic temperature rise of cement and hydration heat coefficient and For example, based on engineering experience and combined with the trend of the temperature curve of the monitoring point, the value range is determined as shown in Table 1.

[0104] Table 1 Thermal parameters to be inverted and their value ranges

[0105]

[0106] Then, Latin hypercube sampling method is used to sample from the value range to obtain multiple sets of design parameter values ​​of the thermal parameters to be inverted.

[0107] Latin Hypercube Sampling (LHS) is an efficient stratified sampling technique for generating approximately random sample points in a multidimensional parameter space. In the LHS method, the value range of each parameter is divided into several intervals of equal probability, a random value is selected in each sub-interval, and then these values ​​are combined into samples. This method achieves complete coverage of the parameter range by ensuring that the marginal distribution of each parameter is evenly stratified. The LHS method can obtain a sample set representing the entire parameter space with a small number of samples, thereby improving sampling efficiency. Therefore, the embodiment of the present invention uses the LHS method to extract parameter combination samples. Exemplarily, multiple sets of design parameter values ​​of thermal parameters to be inverted obtained using LHS are shown in Table 2.

[0108] Table 2 Design parameter values ​​of thermal parameters to be inverted

[0109]

[0110] Furthermore, the embodiment of the present invention inputs multiple sets of design parameter values ​​into the original temperature mechanism model to obtain corresponding design temperature values.

[0111] Based on the heat conduction theory of concrete, the "MODEL CHANGE" feature in ABAQUS software is used to simulate the layer-by-layer pouring process of concrete. In this process, all units are initially set to the "dead" state, and then the corresponding pouring layer units are gradually "activated" according to the actual order of concrete pouring. In addition, the UMATHT subroutine is secondary developed using the FORTRAN language to simulate the heat generation during cement hydration. At the same time, the FILM subroutine is used to set the heat transfer coefficient and the ambient temperature. Then, each set of design parameter values ​​in Table 2 is input into formula (1) in ABAQUS software to obtain the design temperature value corresponding to each set of design parameter values.

[0112] Step b: construct an inversion proxy model consisting of a neural network model and a regression prediction model.

[0113] Due to its good material properties, mass concrete is widely used in construction projects. However, due to the influence of test conditions, construction technology, construction environment and other factors, the design parameter values ​​of the thermal parameters to be inverted selected in the design stage have certain errors, which fail to fully, accurately and truly reflect the cement hydration heat process in the construction stage, resulting in deviations in the numerical simulation calculation results of the temperature field, which may even deviate from the actual situation. In order to improve the accuracy of the prediction of the temperature field of mass concrete, the embodiment of the present invention constructs an inversion proxy model based on measured temperature data and a temperature mechanism model to mine the complex nonlinear relationship between the thermal parameters to be inverted and the temperature field of mass concrete to invert and optimize the thermal parameters, thereby providing data support for the construction of subsequent fusion models.

[0114] The inversion proxy model used in the embodiment of the present invention is a Stacking ensemble model. The Stacking ensemble method is an ensemble method using a hierarchical structure, combining Boosting and Bagging. The principle is to select multiple algorithm models as base learners to mine and analyze the original data, and then output the prediction results to the meta learner for fitting.

[0115] The base learners in the Stacking ensemble model of the embodiment of the present invention are a neural network model and a regression prediction model; the meta learner in the Stacking ensemble model is a linear regression model. The embodiment of the present invention uses the neural network model to effectively handle complex nonlinear problems, and at the same time uses the robustness of the regression prediction model to improve the stability and generalization ability of the Stacking ensemble model.

[0116] The above neural network model is preferably a BP neural network model.

[0117] The advantage of BP neural network lies in its powerful nonlinear mapping ability and self-learning ability. It adjusts the network weights through the back propagation algorithm, so as to approximate any complex nonlinear relationship. However, the traditional BP neural network model also has some limitations, such as easy to fall into the local minimum, slow convergence speed, fixed learning rate and short learning memory cycle. In order to improve the accuracy, computational efficiency and stability of the algorithm, the embodiment of the present invention optimizes and improves the Sigmoid activation function of the BP neural network model.

[0118] The Sigmoid function is an S-shaped curve. There are saturation regions at both ends of its domain. In this region, the output of the network node is slow to respond to the weight adjustment. As a result, even if the network error is large, it is difficult to quickly reduce the error by adjusting the weight, thus falling into the flat area of ​​the error surface, affecting the convergence speed of the network. To solve this problem, the embodiment of the present invention uses the steepness factor , to adjust the shape of the Sigmoid function and expand the non-saturated region, as shown in formula (9). Before training begins, the steepness factor Initialized to 1, if the error value is still large but the error change is small during training, you can increase it appropriately value, such as 1.1, 1.2, etc., to leave the flat area; once you leave, you can The value is restored to the initial value. The improved method can significantly reduce the number of network iterations and improve training efficiency.

[0119] (9)

[0120] In the formula, is the input value, is the output value, is the steepness factor.

[0121] The regression prediction model in the embodiment of the present invention is a correlation vector regression model with multiple target outputs constructed by combining a correlation vector regression model and a regression chain.

[0122] Relevance Vector Machine Regression (RVR) is a model that applies Bayesian methods to statistical learning. It overcomes the shortcomings of the Support Vector Regression (SVR) model: (1) the kernel function selection must satisfy the Mercer condition; (2) the SVM prediction results lack probabilistic expression; (3) the SVM model parameters are difficult to determine; (4) the sparsity is insufficient, etc. It also inherits the advantages of SVR in dealing with small samples, nonlinear and high-dimensional problems.

[0123] The relevant vector regression model can be expressed as follows: Assume that there is a training data set of ,in is the input vector, is the corresponding output, and both the input and output are independently distributed samples, and it is considered that for Regression function on input The noise output is:

[0124] (10)

[0125] In the formula The mean is 0 and the variance is Gaussian noise; Is the serial number.

[0126] RVR only supports a single output, but considering that the three parameters in the adiabatic temperature rise inversion are closely related to each other, it is necessary to expand RVR into a prediction model that can handle multi-target outputs. RegressorChain is a method for solving multi-output regression problems. Its principle is to create a model for each output and use the output of the previous model as input to enhance the prediction of the next model during prediction, so as to capture the correlation between different output targets. When there is a dependency between output variables, using regression chains can provide better prediction performance, so RegressorChain is combined with RVR to construct a correlation vector regression prediction model (Multiple-RVR, Multi-RVR) for multi-target outputs.

[0127] The type of kernel function is a factor that has a greater impact on the RVR prediction accuracy, which determines the distribution of sample data mapped to high-dimensional space. Common kernel functions include linear kernel function, polynomial kernel function and radial basic function (RBF). Considering the prediction accuracy and convergence speed, the embodiment of the present invention selects the radial basis function as the kernel function of RVR.

[0128] The meta-learner used in the embodiment of the present invention is a linear regression (LR) model, which is a simple linear model and can avoid overfitting.

[0129] Step c: taking the design temperature value as input and the design parameter value as output, training the inversion agent model, and obtaining the mapping relationship between the temperature and thermal parameters of the mass concrete.

[0130] In the embodiment of the present invention, the design parameter values ​​and their corresponding design temperature values ​​are taken as a group to obtain multiple groups of values, which are then divided into a training set Train_set and a test set Test_set in a ratio of 8:2 to train and verify the inversion proxy model, and obtain the mapping relationship between the temperature of the mass concrete and the thermal parameters to be inverted.

[0131] During the training process, 5-fold cross validation was used to avoid overlapping of the training samples of the base learner and the meta-learner. In each cross validation, the training data was used to train the model, and the prediction results of the validation set and the test set were obtained. Subsequently, these prediction results were used as the training set and test set of the meta-learner, and then linear regression was used as the meta-learner to continue training. After the training was completed, the mapping relationship between the temperature of the mass concrete and the thermal parameters to be inverted was obtained.

[0132] Step A2: determining the inversion value of the thermal parameter according to the actual temperature measurement value and the mapping relationship, and substituting the inversion value into the original temperature mechanism model to obtain the temperature mechanism model.

[0133] In the embodiment of the present invention, the measured temperature value is input into the mapping relationship to obtain the corresponding total adiabatic temperature rise of cement. The inversion value and hydration heat coefficient The inversion value and hydration heat coefficient The inversion value is then substituted into the original temperature mechanism model to obtain the temperature mechanism model.

[0134] The inversion values ​​of the total adiabatic temperature rise of cement and the hydration heat coefficient in the adiabatic temperature rise equation determined by the mass concrete temperature prediction method of the present invention can comprehensively, accurately and truly reflect the cement hydration heat process in the construction stage, ensuring that the predicted temperature of the mass concrete determined by it is more accurate, thereby facilitating the adoption of corresponding temperature control measures to ensure the construction quality of the mass concrete.

[0135] The second aspect of the present invention discloses a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the mass concrete temperature prediction method are implemented.

[0136] The effectiveness of the method of the present invention will be verified by more specific examples below.

[0137] (1) Dataset extraction

[0138] There are many factors that affect the temperature change process during the concrete construction period, such as cement mix ratio, location information of monitoring points, concrete bulk density, specific heat, thermal conductivity, concrete heat dissipation coefficient to air (related to wind speed), ambient temperature and humidity, and cooling water flow rate. Based on the collected construction information, the factors affecting the temperature of large-volume concrete are classified into: location information, environmental information, water cooling information, concrete material information, thermal parameters and time information. The detailed influencing factors are shown in Table 3.

[0139] Table 3 Classification of factors affecting concrete temperature

[0140]

[0141] In the machine learning model, all input variables need to be converted into a time-related sequence format in order to match the temperature time series output by the machine learning model. Static information such as design information, cooling water pipe layout, and concrete material properties remain unchanged at each time point. Dynamic information such as water cooling data and air temperature are input according to the temperature change law during the actual measurement and construction period. arrive The sensors are numbered in the order of , and their maximum information coefficient with concrete temperature is calculated to evaluate the correlation. The correlation relationship is shown in Table 4. By comparison, it is found that the MIC value of the water temperature at the outlet is the largest, indicating that it has the strongest correlation with the temperature of the mass concrete; while the MIC values ​​of humidity and wind speed are both lower than 0.3, indicating that they have a weak correlation with the temperature of the mass concrete, which may be caused by the distance between the sensor location and the concrete pouring site. Based on the correlation analysis results of the temperature influencing factors in Table 4, 14 influencing factors other than humidity and wind speed are selected as the key features of the data model for data extraction, and the data set is constructed accordingly.

[0142] Table 4 Correlation analysis of temperature influencing factors

[0143]

[0144] (2) Comparison of prediction model performance

[0145] Python was used to build error compensation data models based on improved BP neural network, SVR, and XGBoost respectively; the 447 groups of data obtained were divided into training sets and test sets according to 8:2. After data preparation was completed, BAS and the improved beetle whisker search algorithm (IBAS) of the embodiment of the present invention were used to optimize the hyperparameters in XGBoost respectively, and the ranges of max_depth, learning_rate, and n_estimators were set to [2,10], [0.01,1], [2,100] respectively, the maximum number of iterations was initialized to 200, and the number of beetle groups of IBAS was selected to be 20, 30, and 40 for comparative analysis.

[0146] The iterative optimization curves of BAS and IBAS search algorithms are as follows: Figure 3 As shown in the figure, it can be seen that by increasing the number of longhorn beetles, judging "continuous stability" and using variable step length, the search rate and accuracy of the original BAS are significantly accelerated. The original BAS is in a stagnant state during the iteration process, and it continues until the maximum number of iterations, that is, it falls into the local optimal solution; on the other hand, IBAS uses a decaying exponential function to update the step length, avoids falling into the local optimal solution in the early stage of iteration, and fully searches the global and local space to obtain the optimal fitness value. In the later stage of iteration, through the "continuous stability" judgment, the step length is increased to expand the search space and improve the search efficiency. Overall, the IBAS algorithm is better than the original BAS algorithm, which proves the rationality and superiority of the proposed IBAS algorithm. As the number of longhorn beetle groups increases, the fitness value of the IBAS algorithm decreases. Therefore, the number of longhorn beetle groups is selected as 20 to perform hyperparameter combination optimization on the XGBoost model.

[0147] According to the iterative optimization results of BAS and IBAS, the max_depth, learning_rate and n_estimators of the BAS-XGBoost model are set to 6, 0.6501 and 25, respectively, and the max_depth, learning_rate and n_estimators of the IBAS-XGBoost model are set to 4, 0.6170 and 26, respectively; the parameter settings of the improved BP neural network are shown in Table 5; the SVR model uses the rbf kernel function; the max_depth, learning_rate and n_estimators of the XGBoost model without hyperparameter optimization are set to 6, 0.5 and 20 respectively according to experience.

[0148] Table 5 Initial parameter settings of different models

[0149]

[0150] Comparison of the predicted results and actual values ​​of the error compensation values ​​of the mechanism models of the five models Figures 4 to 8 As shown, it can be seen that the IBAS-XGBoost evaluation model established in the embodiment of the present invention has consistent prediction performance with the improved BP neural network and SVR model, and the accuracy is relatively high; although there are large residuals in individual test samples, the variation range of the overall data residuals is small. It can be clearly seen that after using IBAS to optimize hyperparameters, the prediction effect of the XGBoost model has been greatly improved, the residual has become smaller, and the accuracy of the error compensation value prediction has been further improved; and BAS is trapped in the local optimal solution and cannot obtain the best hyperparameter combination, resulting in the prediction effect of the BAS-XGBoost model is not as good as the XGBoost model set according to experience.

[0151] The performance comparison results of the error compensation data model built based on IBAS-XGBoost and other commonly used models are shown in Table 6. It can be seen that the SVR model has the worst performance, and its correlation coefficient Only 0.5490, while the other four models The IBAS-XGBoost model showed the best results compared with the BP neural network. It is 16% higher, and both MSE and MAE are reduced by more than 30%, with the highest and the lowest MSE and MAE, indicating that IBAS-XGBoost is better than the other four models in terms of accuracy and has strong representativeness and superiority. Because the BAS algorithm falls into a local optimal solution and fails to obtain the optimal hyperparameter combination, BAS-XGBoost , MSE, and MAE are the lowest values ​​among similar XGBoost models.

[0152] Table 6 Comparison of performance indicators of different model test sets

[0153]

[0154] Furthermore, based on the temperature mechanism model constructed according to the inversion analysis results of the adiabatic temperature rise parameters and the error compensation data model constructed based on the above content, the temperature prediction results of the fusion model for different monitoring points are obtained as follows: Figures 9 to 14 As shown in the figure, it can be seen that the data model effectively compensates for the error of the output results of the temperature mechanism model by considering the uncertain temperature influencing factors during the construction of large-volume concrete, so that the fusion model can more accurately predict the concrete temperature and its change process. Since the mean absolute percentage error (MAPE) can measure the relative size of the deviation between the monitoring data and the predicted results, MAPE is used as an indicator for evaluation.

[0155] The MAPE of different monitoring points is shown in Table 7. Taking point A on the waterfront as an example for analysis, the temperature monitoring value is basically flat from the fifth to the sixth day, but the prediction result of the temperature mechanism model has been declining, and after the fifth day, there is a large deviation from the monitoring data, with a MAPE of 4.58%; while the fusion model corrects the output result of the mechanism model through the machine learning model, and the time of the highest temperature is almost consistent with the actual situation, and the overall change trend is basically consistent with the monitoring data, with a MAPE of 1.16%, which is 74.7% lower than that of the mechanism model. It can be observed from Table 7 that after the error compensation of the prediction result of the temperature mechanism model by the machine learning model, the accuracy of the temperature prediction of each monitoring point is significantly improved, and the MAPE is reduced by more than 70%, which proves that the fusion model proposed in the embodiment of the present invention is effective and feasible for the prediction of temperature changes during the construction of large-volume concrete. Compared with the single temperature mechanism model, it more comprehensively considers various factors in the construction process, and avoids the problem of low accuracy of the data model during the construction process due to insufficient data during the construction period, thereby improving the accuracy and reliability of temperature prediction.

[0156] Table 7 Comparison of MAPE of different models

[0157]

[0158] The embodiment of the present invention also verifies the effect of thermal parameter inversion.

[0159] A case analysis is conducted by taking the arc semi-gravity retaining wall on the left bank of a reservoir inlet as an example.

[0160] The semi-gravity retaining wall finite element model was established, with a total of 660,652 units and 134,548 nodes in the dam body. , the calculated heat release coefficient is 63.58. The annual average temperature is 16℃. The water-cement ratio of concrete is 0.40, the sand ratio is 37, and the 28-day compressive strength is 35.9MPa. The mix ratio parameters are shown in Table 8. Take the correction coefficient , the thermal performance index parameters of concrete are calculated using the percentage weighted method. The thermal performance parameters of concrete calculated in this way are shown in Table 9.

[0161] Table 8 Concrete mix parameters

[0162]

[0163] Table 9 Thermal performance parameters of concrete

[0164]

[0165] According to formula (1), select the total adiabatic temperature rise of cement , heat of hydration constant and As the thermal parameters to be inverted, the value range is determined based on engineering experience and combined with the trend of the temperature curve of the monitoring point as shown in Table 1 above. The LHS method is used to extract parameter combination samples, and a total of 50 sets of samples are extracted as shown in Table 2.

[0166] Parameter inversion calculation was performed on the first layer of cast-in-place where cooling water pipes were laid. The simulation was a monitoring analysis of the retaining wall from the beginning of the first layer casting to the 15th day. By using samples with different adiabatic temperature rise parameters, the time and value of the maximum temperature inside the concrete were basically different. Taking the sample {48.48, 1.40, 1.55} as an example, its maximum temperature was 58.22°C, which occurred 52.25h after the casting was completed.

[0167] After the Stacking integrated model is trained by constructing a training set with the extracted samples and obtaining the mapping relationship between the temperature of the mass concrete and the thermal parameters to be inverted, the measured temperature value is input into the Stacking integrated model to obtain The inversion values ​​of the three parameters are input into the established finite element model, and the predicted temperature value at the monitoring point is output, which is then compared with the measured value for verification. Monitoring points A, B and C are all located on the first layer of the first section of the retaining wall, the distance between monitoring point A and the foundation is 0.1m; the distance between monitoring point B and the foundation is 1.5m; and the distance between monitoring point C and the foundation is 2.9m. Point B, located at the core position, is selected as the feature point. The measured temperature value of monitoring point B is input into the Stacking integrated model, and the inversion value of the adiabatic temperature rise parameter is obtained as shown in Table 10. The inverted parameter value is substituted into formula (1), and the temperature of point A on the water surface is re-simulated. The simulation calculation value corresponding to the inversion value output is closer to the temperature monitoring value in the actual construction environment. Fig.15 As shown, it is feasible to use the Stacking integrated model to invert the parameters of the adiabatic temperature rise equation. Fig.15 The test value output is the output corresponding to the design parameter value in the embodiment of the present invention.

[0168] Table 10 Inversion values ​​of adiabatic temperature rise parameters

[0169]

[0170] Furthermore, in order to evaluate the performance of the constructed Stacking integrated model, the single models (improved BP neural network, Multi-RVR model, SVR model) were used as controls, and the correlation coefficients were selected. , mean square error and mean absolute error As a performance evaluation indicator. The correlation coefficient indicates the degree to which the regression model fits the data, and can be used to compare different models. The closer its value is to 1, the better the model fits. The mean square error indicates the mean of the sum of squares of the difference between the predicted value and the actual value, reflecting the average difference between the predicted value and the actual value. The smaller its value is, the better the model prediction effect is. The mean absolute error indicates the average of the absolute value of the difference between the predicted value and the actual value, reflecting the average deviation between the predicted value and the actual value. When the predicted value is exactly the same as the actual value, it is 0. The larger the deviation, the larger the value.

[0171] Since the SVR model is also a single-target output regression algorithm, the same processing method as RVR is used to transform it into a multi-target output regression prediction (Multi-SVR) using RegressorChain. Based on Python language, Scikit-learn library and Keras framework, inversion proxy models based on improved BP neural network, Multi-RVR, Multi-SVR and Stacking integrated model are built respectively; the train_test_split method in the scikit-learn library is used to divide the 50 groups of inversion sample data into training set and test set. The initial parameter settings of each inversion proxy model are shown in Table 5 above.

[0172] The predicted value of the inversion surrogate model for the test set is compared with the sample design value. Figures 16 to 18 As shown, it can be seen that the prediction effect of the Multi-SVR model is the worst, especially in the parameter and Its prediction value is basically in an unchanged state, while Multi-RVR has better regression prediction effect than SVR due to its sparsity and the ability to automatically update hyperparameters. The Stacking integrated model established by the present invention is consistent with the prediction performance of the improved BP neural network, Multi-RVR, and Multi-SVR. It can maintain basically the same trend as the sample design value in the three inversion parameters, and it can be seen that after combining different types of algorithms using the Stacking integration method, the prediction results of the model are greatly improved, the residual becomes smaller, and the prediction accuracy is further improved. For example, in the second group of samples, the Stacking integrated model makes the absolute error closer to 0 by fusing the prediction results of the improved BP neural network and Multi-RVR.

[0173] The performance index comparison results of the improved BP neural network, Multi-RVR, Multi-SVR and Stacking integrated model are shown in Table 11. It can be seen that except for Multi-SVR, the accuracy of the mean square error and mean absolute error of the predicted values ​​and the designed values ​​of the other three models are at the same level, and the correlation coefficient are all greater than 0.9. In general, when When it is greater than 0.8, the model is considered effective. Therefore, it can be considered that except for Multi-SVR, the other models can make accurate predictions. Compared with the improved BP neural network and Multi-RVR, the Stacking ensemble model They are 3.45% and 1.35% higher respectively, the mean square error is reduced by 1.31% and 0.78% respectively, and the mean absolute error is reduced by 3.97% and 2.90% respectively, which shows the effectiveness of the model proposed in the present invention. It also shows that the Stacking ensemble model is superior to the other four models in accuracy and is more representative and superior.

[0174] Table 11 Comparison of performance indicators of different model test sets

[0175]

[0176] The inversion values ​​of the total adiabatic temperature rise of cement and the hydration heat coefficient in the adiabatic temperature rise equation determined by the mass concrete temperature prediction method of the present invention can comprehensively, accurately and truly reflect the cement hydration heat process in the construction stage, ensuring that the predicted temperature of the mass concrete determined by it is more accurate, thereby facilitating the adoption of corresponding temperature control measures to ensure the construction quality of the mass concrete.

[0177] During the construction of large-volume concrete, uncertain factors such as environmental conditions and meteorological information will cause deviations in the temperature mechanism model even after parameter inversion optimization; and the single data model lacks an in-depth understanding of the hydration heat reaction of large-volume concrete, and the limited amount of data makes it impossible to build a high-accuracy temperature prediction model. Therefore, the present invention proposes a parallel fusion model, which is based on the prediction results of the temperature mechanism model and uses a machine learning model that takes uncertain factors into account to correct local errors, so that the resulting temperature prediction value is more accurate.

[0178] The above are only several embodiments of the present invention and are not intended to limit the present invention in any form. Although the present invention is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any technician familiar with the profession, without departing from the scope of the technical solution of the present invention, using the above disclosed technical content to make slight changes or modifications are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A method for predicting temperature of mass concrete, characterized in that: include: Step 1: obtaining the temperature prediction value of each measuring point of the mass concrete predicted by the temperature mechanism model, and determining the temperature error between the temperature prediction value and the corresponding temperature measured value; Step 2: training a machine learning model based on the temperature influencing factor data of mass concrete and the temperature error; Step 3: Input the measured temperature influencing factor data into the trained machine learning model to obtain the error compensation prediction value; Step 4: Correcting the temperature prediction value using the error compensation prediction value; The step 1 also includes: Step A, performing parameter inversion on the thermal parameters in the original temperature mechanism model to obtain the temperature mechanism model, specifically comprising: Step A1, determining the mapping relationship between the temperature of the mass concrete and the thermal parameters, specifically comprising: obtaining multiple sets of design parameter values ​​and corresponding design temperature values ​​of the thermal parameters to be inverted in the original temperature mechanism model; constructing an inversion proxy model composed of a neural network model and a regression prediction model; taking the design temperature value as input and the design parameter value as output, training the inversion proxy model, and obtaining the mapping relationship between the temperature of the mass concrete and the thermal parameters; The inversion proxy model is a Stacking integration model; the base learners in the Stacking integration model are a neural network model and a regression prediction model; the meta learner in the Stacking integration model is a linear regression model; the regression prediction model is a correlation vector regression model of multi-target output constructed by combining a correlation vector regression model and a regression chain; Step A2: determining the inversion value of the thermal parameter according to the actual temperature value and the mapping relationship, and substituting the inversion value into the original temperature mechanism model to obtain the temperature mechanism model.

2. The method for predicting temperature of mass concrete according to claim 1, characterized in that: The machine learning model is an XGBoost model; Then the step 2 is specifically as follows: The temperature influencing factor data of the mass concrete is used as input, and the temperature error is used as output to train the XGBoost model.

3. The method for predicting temperature of mass concrete according to claim 2, characterized in that: The XGBoost model includes multiple hyperparameters; Optimizing the multiple hyperparameters using a longicorn search algorithm; The fitness function in the longicorn beetle whisker search algorithm is determined according to the correlation coefficient and mean square error between the temperature prediction value and the corresponding temperature measured value.

4. The method for predicting temperature of mass concrete according to claim 3, characterized in that: The step size adjustment method of the longicorn beard search algorithm in the iterative process is: Step S1: Obtain the fitness value in the process of adjusting the step length using the step length adjustment formula at the preset number of iterations n Standard deviation within ; Step S2: When the standard deviation is less than or equal to the preset threshold, the number of iterations is increased by 1. n ≥ Maximum number of iterations m When the step size is expanded to a preset multiple of the step size corresponding to the current number of iterations, the number of iterations is n When the standard deviation is greater than a preset threshold or n < m When , the step length adjustment formula is used to adjust the step length; Repeat steps S1 and S2 until the maximum number of iterations is met or the preset accuracy requirement is met.

5. The method for predicting temperature of mass concrete according to claim 1, characterized in that: The neural network model is a BP neural network model; The activation function of the BP neural network model includes a steepness factor, and the steepness factor is ≥1.

6. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the mass concrete temperature prediction method according to any one of claims 1 to 5 are implemented.

Images