Shield tunneling carbon emission prediction method and system based on LSTM and bitter fish optimization

Abstract

The invention provides a shield tunneling carbon emission prediction method and system based on LSTM and bitter fish optimization, and the method comprises the steps: obtaining a segment ring number, stratum parameters, construction power consumption and an actual carbon emission value, and taking the values as a training set; establishing an initial LSTM model, and constraining the structural parameter range of the initial LSTM model; optimizing the structural parameters of the LSTM model by using a BFO method to obtain an LSTM model with optimal training parameters; the training set is adopted to train the LSTM model with the optimal training parameters, effect evaluation is carried out, and an LSTM model for prediction is obtained; and inputting the number of segment rings, the stratum parameters and the construction power consumption in a period of time before the to-be-predicted moment into the LSTM model for prediction to obtain a predicted carbon emission value corresponding to the to-be-predicted moment, and performing reverse normalization processing to obtain the carbon emission of an original scale. According to the method, the accuracy and precision of carbon emission in shield tunneling of the highway tunnel can be improved, and the scheme comparison and selection time of the highway tunnel based on carbon emission is saved.

Description

Technical Field

[0001] The present invention relates to the technical field of carbon emission, and in particular to a shield tunneling carbon emission prediction method and system based on LSTM and bitterfish optimization algorithm. Background Art

[0002] As the concept of carbon peak and carbon neutrality gradually deepens in the highway field, carbon quantification of a large number of sub-projects in highway projects has gradually become an important topic. Tunnels are areas with high concentrations of carbon dioxide. In recent years, my country's transportation industry has experienced rapid development, especially infrastructure projects such as highways and high-speed rail projects. This has not only greatly promoted the integration of regional economies, but also put forward higher requirements for infrastructure construction.

[0003] The implementation of infrastructure projects is often accompanied by a large number of tunnel projects. During the entire life cycle of these projects, from material transportation to construction and later maintenance, each link involves energy consumption and carbon dioxide emissions. As a costly link in tunnel construction, shield machine excavation requires careful planning in the early stages and full consideration of the carbon emissions of different schemes. Therefore, the refined management and prediction of carbon emissions for excavation schemes becomes particularly prominent.

[0004] Scientific and rational analysis of carbon emissions during shield tunneling is crucial to promoting the green development of tunnel engineering. However, the refined accounting process of carbon emissions is not only technically demanding, but also very complex, and traditional accounting methods are difficult to achieve the required level of speed and refinement. Although tunnel engineering, as an important large-scale infrastructure, plays a key role in shortening traffic distances and creating underground space, its huge consumption of materials and energy during construction has also made it a major source of carbon emissions.

[0005] As the country attaches great importance to carbon emission accounting and control, it has become a trend to include large-scale transportation infrastructure projects in the national carbon emission accounting system. Carbon emission measurement is not only the basis for engineering carbon emission assessment, but also the key to achieving carbon emission trading, control, and energy conservation and emission reduction effect evaluation. Therefore, the development of carbon emission measurement methods suitable for tunnel engineering construction is of great significance to promoting the green development of engineering.

[0006] LSTM (Long Short-Term Memory Network), as a widely used RNN (Recurrent Neural Network), is known for its excellent memory capacity and robustness, and can adapt to different strata and construction conditions. LSTM is often used in existing technologies to effectively predict carbon emissions in tunnel projects. However, the prediction effect of LSTM is significantly affected by its structural parameters, including the length of the time series, the number of units in the hidden layer, and the number of layers in the hidden layer. If the optimal structural parameters cannot be used, the prediction effect will be greatly reduced. Summary of the invention

[0007] In view of the defects in the prior art, the purpose of the present invention is to provide a shield tunneling carbon emission prediction method and system based on LSTM and bitter fish optimization.

[0008] According to one aspect of the present invention, a shield tunneling carbon emission prediction method based on LSTM and bitterfish optimization is provided, comprising:

[0009] The number of segment rings, formation parameters, construction power consumption, and actual carbon emissions are obtained as training sets;

[0010] Establishing an initial LSTM model and constraining the range of structural parameters of the initial LSTM model;

[0011] The structural parameters of the LSTM model are optimized using the BFO method to obtain an LSTM model with optimal training parameters;

[0012] Using the training set, training the LSTM model with the optimal training parameters, and performing effect evaluation to obtain an LSTM model for prediction;

[0013] The number of segment rings, formation parameters, and construction power consumption in a period of time before the predicted time are input into the prediction LSTM model to obtain the corresponding predicted carbon emission value at the predicted time, and then denormalized to obtain the carbon emission value at the original scale.

[0014] Preferably, the acquisition of segment ring numbers, formation parameters, construction power consumption and actual carbon emission values ​​as a training set includes:

[0015] Obtain the number of segments in a period of time before the prediction time, as well as the corresponding formation parameters and construction power consumption;

[0016] Based on the construction power consumption, the actual carbon emission value corresponding to each ring of the segment is obtained by converting the carbon emission coefficient;

[0017] The segment ring number, formation parameters, construction power consumption and actual carbon emission values ​​are preprocessed as follows:

[0018] Eliminate the data group corresponding to construction power consumption of 0;

[0019] Use box plots to process outliers in the selected time period and remove data groups containing outliers;

[0020] The minimum and maximum normalization methods were used to normalize the eliminated data, and the formation parameters, construction power consumption, and actual carbon emission values ​​corresponding to the number of preprocessed segment rings were obtained.

[0021] Preferably, the formation parameters include: cohesion parameters, internal friction angle parameters, penetration resistance parameters, earth pressure parameters and tunnel burial depth parameters of the formation where the segments of the tunnel to be constructed are located.

[0022] Preferably, establishing an initial LSTM model comprises:

[0023] Add three neural network layers based on the RNN neural network, namely: forget gate, input gate, and output gate, to obtain the initial LSTM model;

[0024] The initial LSTM model is specifically:

[0025] f t =σ(W f ·[h t-1, x t ]+b f )

[0026] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0027]

[0028] o t =σ(W o [h t-1 ,x t ]+b o )

[0029] h t =o t tanh(C t )

[0030] Among them, C t-1 is the long-term memory, i.e. the output from the previous unit module, f t is the forgetting factor, h t-1 is short-term memory, W is the weight matrix of the gate, b is the bias term, x t is the event information, the subscript f represents the forget gate, the subscripts i and c represent the first and second parts of the input gate respectively, o represents the output gate, and i t is the long-term memory output from the forget gate, is the short-term memory output from the learning gate, W f is the weight matrix of the forget gate, [h t-1 ,x t ] means connecting two vectors into a longer vector, b fis the bias term of the forget gate, σ is the sigmoid function, W f is the weight matrix of the forget gate, W i is the weight matrix of the input gate, b i is the bias term of the input gate.

[0031] Preferably, the structural parameters of the initial LSTM model include:

[0032] The feature dimension input_size of the input sequence, which specifies the size of the last dimension of the input tensor;

[0033] The size of the hidden state hidden_size, that is, the output dimension of the LSTM unit, which determines the size of the hidden state in the LSTM unit;

[0034] The number of LSTM layers num_layers, the default is 1;

[0035] Whether to use the bias term bias, the default is True;

[0036] Whether the first dimension of the input and output tensors is the batch size batch_first, the default is False;

[0037] Whether to perform dropout operation. The default value is 0, which means no dropout operation is applied.

[0038] Whether to use bidirectional LSTMbidirectional, the default is False.

[0039] Preferably, constraining the structural parameter range of the initial LSTM model includes:

[0040] The feature dimension input_size of the input sequence is set in the range of 8-12;

[0041] The size of the hidden state hidden_size is set in the range of 16-64;

[0042] The number of LSTM layers num_layers is set in the range of 1-6;

[0043] Whether to use the bias term bias is True;

[0044] Whether to discard the operation dropout is set to 0.3.

[0045] Preferably, the use of the BFO method to optimize the structural parameters of the LSTM model to obtain an LSTM model with optimal training parameters includes:

[0046] Iterative solution:

[0047] The parameters of LSTM are used as the initial solution in the optimization of BFO method, and each solution represents a bitter fish;

[0048] Evaluating the fitness of each solution using an objective function;

[0049] Each solution searches for and attempts to occupy the optimal solution during the iteration process, and updates the position by randomly selecting and competing with the current solution;

[0050] If the current solution fails to successfully occupy the optimal solution, the current solution does not need to randomly search for a new position;

[0051] If the optimal solution is occupied, the current solution is spawned, that is, a new solution is generated around the current solution;

[0052] Evaluate the fitness of each solution and remove solutions that do not meet a set threshold;

[0053] Update parameters:

[0054] During the iteration process, the parameters of the BFO method are updated, including: initial population size ips, dimension Dim, initial step size J, initial probability parameter P, exponent a of the decay rate, initial parameter U, initial radius R at spawning, and maximum number of iterations it max , number of test repetitions rp;

[0055] Iteration termination:

[0056] When the maximum number of iterations is reached, the iteration stops and the currently found optimal solution is output.

[0057] Preferably, the effect evaluation of the trained LSTM model includes:

[0058] Calculate the mean square error MSE and sample regression value R between the model prediction value and the true value 2 , as an evaluation indicator of model prediction effect:

[0059]

[0060] Where: N is the number of samples, y i is the predicted value of shield target carbon emissions, is the actual value of the shield tunneling target carbon emissions, is the average value of the actual value of the shield target carbon emissions; the smaller the MSE, the higher the R 2 The larger it is, the higher the fitting accuracy of the LSTM prediction model is and the better the prediction performance is.

[0061] Preferably, the carbon emission prediction result is subjected to a denormalization process to obtain the carbon emission in the original scale, including:

[0062] x=x'×(xmax -x min )+x min

[0063] Among them: x is the original data, x min is the minimum value of a type of data in the sample; x max is the maximum value of the same type of data in the sample, and x' is the normalized data.

[0064] According to a second aspect of the present invention, a shield tunneling carbon emission prediction system based on LSTM and bitterfish optimization is provided, comprising:

[0065] Data module: obtain the number of segment rings, formation parameters, construction power consumption and actual carbon emissions as training sets;

[0066] Initial modeling module: establish an initial LSTM model and constrain the structural parameter range of the initial LSTM model;

[0067] BFO optimization module: using the BFO method to optimize the structural parameters of the LSTM model to obtain an LSTM model with optimal training parameters;

[0068] Training and evaluation module: using the training set to train the LSTM model with the optimal training parameters, and performing effect evaluation to obtain the LSTM model for prediction;

[0069] Application module: Input the number of segment rings, formation parameters, and construction power consumption in a period of time before the predicted time into the LSTM model used for prediction, obtain the corresponding predicted carbon emission value at the predicted time, and perform inverse normalization processing to obtain the carbon emission value at the original scale.

[0070] Compared with the prior art, the embodiments of the present invention have at least one of the following beneficial effects:

[0071] The shield tunneling carbon emission prediction method and system based on LSTM and bitter fish optimization in the embodiment of the present invention can improve the accuracy and precision of the quantification of carbon emissions from shield tunneling in highway tunnels, and save the time for comparing schemes based on carbon emissions in highway tunnels. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0073] Figure 1 This is a flow chart of a shield tunneling carbon emission prediction method based on LSTM and bitterfish optimization in one embodiment of the present invention;

[0074] Figure 2A flow chart of data preprocessing steps in a preferred embodiment of the present invention;

[0075] Figure 3 The figure is a flowchart of optimization and prediction of LSTM structural parameters by the bitter fish algorithm in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0076] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several variations and improvements may be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0077] In view of the burden of field data collection and the difficulty in making scientific predictions in the early stages of project planning when using the carbon emission coefficient method for measurement, in one embodiment of the present invention, a shield tunneling carbon emission prediction method based on LSTM and bitterfish optimization is provided, such as Figure 1 As shown, the main process is as follows:

[0078] S100, obtaining the number of segment rings, formation parameters, construction power consumption and actual carbon emission values ​​as a training set;

[0079] S200, establishing an initial LSTM model, and constraining a range of structural parameters of the initial LSTM model;

[0080] S300, using the BFO method to optimize the structural parameters of the LSTM model obtained in S200 to obtain an LSTM model with optimal training parameters;

[0081] S400, using the training set obtained in S100, train the LSTM model with the best training parameters obtained in S300, and evaluate the effect to obtain the LSTM model for prediction;

[0082] S500 inputs the number of segment rings, formation parameters, and construction power consumption in a period of time before the prediction time into the LSTM model trained in S400 for prediction, obtains the corresponding predicted carbon emission value at the prediction time, and performs denormalization processing to obtain the carbon emission value at the original scale.

[0083] Through the shield tunneling carbon emissions prediction method in the above embodiment, it is possible to train and predict the carbon emissions of tunnel shield excavation under different working conditions, thereby reducing the complexity of data collection and being able to effectively predict and evaluate the carbon emissions of the project during the design stage.

[0084] In a preferred embodiment of the present invention, S100 is implemented to obtain the number of segment rings, formation parameters, construction power consumption and actual carbon emission values ​​as a training set. Specifically, the following steps can be used:

[0085] S101, obtaining the number of segment rings within a time period before the prediction time, as well as the corresponding formation parameters and construction power consumption;

[0086] The number of segments usually refers to the number of segments spliced ​​into rings during shield tunnel construction. During the shield tunneling process, the shield machine will assemble prefabricated segments (usually concrete or steel) one by one to form a ring structure, which is part of the tunnel. Each completed ring assembly means that the shield machine has moved forward a certain distance. The number of segments is an important indicator of the progress of shield tunnel construction and an important reference for analyzing formation parameters and carbon emissions.

[0087] During the shield tunneling process, each ring of segments will correspond to a set of formation parameters, which may include the cohesion, internal friction angle, penetration resistance, soil pressure and tunnel depth of the formation. These formation parameters can be obtained through geological exploration, formation testing and real-time monitoring of the shield machine's sensors. The power consumption required for each ring of segments will also be recorded.

[0088] The choice of "a time period before the prediction moment" will depend on the actual working conditions of the shield tunneling and the frequency of ground changes. In practical applications, this time period may be several hours or two working shifts of the shield machine to ensure that enough data is collected to accurately reflect the changes in ground conditions. The specific time range needs to be determined according to the specific needs of the project and the actual situation of shield tunneling.

[0089] S102, based on the construction power consumption, the actual carbon emission value corresponding to each ring of the segment is obtained by converting the carbon emission coefficient; specifically, the construction power consumption is multiplied by the carbon emission coefficient to obtain the actual carbon emission value.

[0090] S103, pre-processing the number of segment rings, formation parameters, construction power consumption and actual carbon emission values ​​to obtain data that meets the subsequent model input requirements.

[0091] Furthermore, in order to obtain a better preprocessing effect, in a preferred embodiment of the present invention, data preprocessing can be performed using the following steps: Figure 2 As shown:

[0092] First, remove the data set with original data power consumption of 0 in data0 to obtain data set data1, excluding the impact of shield downtime;

[0093] Then, box plots were drawn using the formation parameters, power consumption, number of segment rings, and actual carbon emissions in data1 to obtain 5 box plots. The groups containing outliers in the box plots were removed to obtain data set data2, eliminating the influence of other outliers.

[0094] Finally, the minimum-maximum normalization method is used to normalize the formation parameters, power consumption, number of segments, and actual carbon emissions in data2 to obtain the data set data3, which is the input and output parameters used to train the LSTM network. The calculation formula for minimum-maximum normalization is as follows:

[0095]

[0096] Among them: x is the original data, x min is the minimum value of a certain type of data in the sample; x max is the maximum value of a certain type of data in the sample, and x' is the normalized data.

[0097] In some specific embodiments, the original data set data0 includes the following data types:

[0098] - Formation parameters: These parameters include the cohesion parameters, internal friction angle parameters, penetration resistance parameters, earth pressure parameters and tunnel burial depth parameters of the formation where the tunnel segments to be constructed are located.

[0099] -Construction power consumption: The construction power consumption corresponding to the set number of segment rings.

[0100] -Actual carbon emission value of each ring segment: the value obtained by multiplying the construction electricity consumption by the carbon emission coefficient.

[0101] Data set data3 is the input and output of the trained LSTM model, as follows:

[0102] -Input data: the number of shield segments, power consumption, and corresponding formation parameters after preprocessing from n moments before the prediction time to one moment before the prediction time. These data are normalized by minimum and maximum values ​​to meet the input requirements of the LSTM model.

[0103] -Output data: Target carbon emissions of shield tunneling at the prediction moment. This is the result predicted by the LSTM model based on the input data, indicating the carbon emissions of shield tunneling at a specific moment. data3 is a preprocessed data set used to train the LSTM model. The input data includes the number of segments, construction power consumption, and formation parameters after normalization, and the output data is the corresponding predicted carbon emissions value. Through training, the LSTM model learns how to predict carbon emissions based on the input data, thereby providing support for the prediction of carbon emissions of shield tunneling in practical applications.

[0104] LSTM (Long Short-Term Memory Network) is a widely used RNN (Recurrent Neural Network) currently known for its excellent memory capacity and robustness, and can adapt to different strata and construction conditions. LSTM is often used in the prior art to effectively predict carbon emissions in tunnel engineering. Therefore, in a preferred embodiment of the present invention, step S200 is implemented to establish an initial long short-term memory network (LSTM) to solve the problem that RNN cannot handle long-distance dependencies. It adds three neural network layers (forget gate, input gate, output gate) to protect the control unit state. The functional relationship between these parameters and memory is:

[0105] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0106] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0107]

[0108] o t =σ(W o [h t-1 ,x t ]+b o )

[0109] h t =o t tanh(C t )

[0110] Among them, C t-1 is the long-term memory (the output from the previous unit module), f t is the forgetting factor, h t-1 is short-term memory, W is the weight matrix of the gate, b is the bias term, x t is the event information, the subscript f represents the forget gate, the subscripts i and c represent the first and second parts of the input gate respectively, o represents the output gate, and i t is the long-term memory output from the forget gate, is the short-term memory output from the learning gate, W f is the weight matrix of the forget gate, [h t-1 ,x t ] means connecting two vectors into a longer vector, b f is the bias term of the forget gate, σ is the sigmoid function, W fis the weight matrix of the forget gate, W i is the weight matrix of the input gate, b i is the bias term of the input gate.

[0111] The initial parameters of the LSTM network are:

[0112] input_size (int): The feature dimension of the input sequence. It specifies the size of the last dimension of the input tensor.

[0113] hidden_size(int): The size of the hidden state, i.e. the output dimension of the LSTM unit. It determines the size of the hidden state in the LSTM unit.

[0114] num_layers (int, optional): The number of LSTM layers, default is 1. If set to a value greater than 1, multiple LSTM layers will be stacked.

[0115] bias (bool, optional): whether to use the bias term, the default is True. If set to False, the LSTM layer will not use the bias term.

[0116] batch_first (bool, optional): Whether the first dimension of the input and output tensors is the batch size, default is False. If set to True, the shape of the input and output tensors should be (batch_size, sequence_length, input_size).

[0117] dropout (float, optional): If non-zero, a dropout operation will be applied on the output of the LSTM layer. It specifies the dropout probability to apply, and defaults to 0, i.e. no dropout operation is applied.

[0118] bidirectional (bool, optional): Whether to use bidirectional LSTM, the default is False. If set to True, a bidirectional LSTM is created that can process the input sequence from the forward and backward directions respectively.

[0119] In order to provide high subsequent optimization efficiency, in a preferred embodiment of the present invention, the initial structure parameters of the LSTM network are restricted, see Table 1.

[0120] Table 1 Initial structure parameter range of LSTM network

[0121]

[0122] The prediction effect of LSTM is significantly affected by its structural parameters, including the length of the time series, the number of units in the hidden layer, and the number of layers in the hidden layer. If the optimal structural parameters cannot be used, the prediction effect will be greatly reduced. Therefore, in a preferred embodiment of the present invention, step S300 is implemented to optimize the structural parameters of the LSTM network using the bitter fish optimization method.

[0123] Bitterling Fish Optimization (BFO) is a nature-inspired optimization algorithm inspired by the life habits of bitterling fish. Bitterling fish have a unique habit of laying eggs in the gill cavity of freshwater mussels to protect the eggs from predators. This survival strategy inspired the design of the Bitterling Fish Optimization algorithm, making it an effective global optimization method.

[0124] Specifically, Figure 3 As shown in the figure, the BFO algorithm uses the parameters of LSTM as the initial population in the optimization algorithm, and updates the parameters through the search strategy of the BFO algorithm, such as finding oysters, oyster competition, mating and escape behavior modeling. In each step, the performance of the current parameters is evaluated according to the fitness function, and the search process is guided by the evaluation results. The structural parameters of the LSTM network can be back-propagated and optimized globally. The proposed method focuses on finding the optimal solution and other suitable solutions, so it is not easy to converge to local optimization. The mean square error (MSE) between the output of the prediction model training set and the true value is iterated until the set error requirement or the maximum number of iterations is reached. Finally, the BFO algorithm converges to a set of optimal or approximately optimal LSTM parameters that can achieve the best performance on specific tasks.

[0125] Furthermore, in a preferred embodiment, embodiment S300, the structural parameters of the LSTM model are optimized using the BFO method to obtain the best judgment value, the best position and the LSTM model for prediction. The specific process is as follows:

[0126] S301, set the parameters of the BFO method: initial population size ips, dimension Dim, initial step size J, initial probability parameter P, exponent a of the decay rate, initial parameter U, initial radius R when spawning, maximum number of iterations it max , the number of test repetitions rp. Table 2 is a parameter setting table of the BFO method in this embodiment.

[0127] Table 2 BFO method parameter setting table

[0128]

[0129] S302, initialize the population: The initial structural parameters of LSTM are used as the initial population (initial solution) in the optimization algorithm. Each solution represents a "bitter fish" or "egg", and these solutions are distributed in the decision space of the problem.

[0130] S303, evaluate solutions: use the objective function to evaluate the fitness of each solution, where the fitness is based on the quality of the solution, such as whether the goal of the problem is minimization or maximization.

[0131] S304, search and occupy oysters (search and select optimal solution): Each fish (solution) searches for and tries to occupy oysters (optimal solution) during the iteration process, and updates the position by random selection and competition with the current solution.

[0132] S305, Escape and Unoccupy the Oyster: If the current solution fails to successfully occupy the oyster, it will escape and randomly search for a new position, which helps the algorithm avoid premature convergence.

[0133] S306, Spawning (Generating New Solutions): Once an oyster is found and occupied, a solution can "spawn", i.e. generate a new solution around the current solution, which helps the algorithm explore the solution space.

[0134] S307, Predator fish (eliminating inferior solutions): Inferior solutions (based on fitness evaluation) may be "predated", i.e. removed from the population, to maintain the quality of the population.

[0135] S308, update parameters: as the iteration proceeds, update the parameters in the algorithm, such as the search step size, to balance the global search and the local search.

[0136] S309, termination condition: When the maximum number of iterations or other termination conditions are reached, the algorithm stops and outputs the currently found optimal solution.

[0137] When BFO is applied to other fields, its processing objects are usually discrete or heterogeneous parameters. In contrast, the parameters of LSTM are usually expressed as continuous values. In addition, the optimization goals in other fields may focus on minimizing costs, maximizing efficiency, or meeting specific performance standards, while the optimization goals of the present invention focus on improving the prediction accuracy of the model, shortening the training time, and enhancing the generalization ability of the model. These differences make the use of BFO to optimize LSTM models a significant challenge and represent an innovation.

[0138] In the innovation process, the input data is first preprocessed, such as normalization, to meet the needs of model input. In addition, the BFO method itself involves the adjustment of multiple parameters, including population size, number of iterations, step size, etc. The selection of these parameters will have a significant impact on the optimization effect. Of course, there are differences in data format between LSTM and BFO, and appropriate fusion and adjustment are required to ensure the consistency and validity of the data.

[0139] It should be noted that the BFO method is specific to LSTM optimization, so the constraints on the range of LSTM model structure parameters help improve the efficiency and effect of the optimization process. The above S301-S309 describes the application of the BFO algorithm in LSTM model structure parameter optimization, which is a specific optimization strategy for LSTM characteristics.

[0140] In the above embodiment, using BFO to optimize the LSTM model can obtain a prediction model with the following advantages:

[0141] Hyperparameter optimization can be achieved: The performance of the LSTM network depends largely on the setting of its hyperparameters, such as learning rate, number of hidden layers and number of units, etc. The BFO algorithm can be used to search for the best combination of these hyperparameters to improve the prediction accuracy of the LSTM model;

[0142] Possessing global search capability: This embodiment can effectively avoid falling into the local optimal solution in a complex parameter space, which is particularly important for finding the optimal structure of the LSTM model;

[0143] High convergence performance: This embodiment can maintain population diversity during the search process by simulating the behavior of bitterlings looking for hosts, thereby improving the convergence speed and stability of the algorithm;

[0144] Strong adaptability: This embodiment can adaptively adjust the search strategy according to the characteristics of the problem, which makes it very flexible and adaptable when dealing with carbon emission prediction problems of different tunnel projects;

[0145] Strong robustness: This embodiment does not rely on specific information of the problem, which enables it to maintain good robustness when facing the uncertainty and variability of carbon emission data in tunnel engineering.

[0146] By applying the BFO method to the hyperparameter optimization of the LSTM model, the accuracy and reliability of tunnel carbon emission prediction can be significantly improved, providing strong decision-making support for environmental protection and energy management of tunnel projects.

[0147] Connecting to the LSTM model with the best training parameters obtained in the above embodiment, in a preferred embodiment, the training data set data3 obtained in step S100 is divided into a training set Train_Set and a test set Valid_Set in a ratio of 7:3, and the optimal solution of the structural parameter group is obtained using S301-S309, and the LSTM network model finally used for prediction is constructed. The number of pre-processed shield segments, power consumption, and corresponding formation parameters from n moments before the prediction moment to one moment before the prediction moment in Train_Set are used as the input of the prediction model, and the target carbon emissions of the shield at the prediction moment are used as the output of the prediction model for training. If the model setting error requirement or the maximum number of iterations is reached, the training is stopped to obtain the prediction model.

[0148] The preprocessed number of shield segments, power consumption, and corresponding formation parameters from n moments before the prediction time to one moment before the prediction time in Valid_Set are used as the input of the prediction model to obtain the model prediction value: the target carbon emissions of the shield at the prediction time.

[0149] Furthermore, in order to estimate the prediction effect, in a preferred embodiment, the mean square error (MSE) between the model prediction value and the true value and the sample regression value (R 2 ), as the evaluation index of model prediction effect, MSE and R 2 The calculation method is as follows:

[0150]

[0151] Where: N is the number of samples, y i is the predicted value of shield target carbon emissions, is the actual value of the shield tunneling target carbon emissions, is the average value of the actual value of the shield target carbon emissions. The smaller the MSE, the higher the R 2 The larger it is, the higher the fitting accuracy of the LSTM prediction model is and the better the prediction performance is.

[0152] Furthermore, in some possible embodiments, a fitness function is also designed, which can directly use MSE or R 2 , or combine the two to form a comprehensive indicator. For example, a fitness function can be designed that minimizes MSE while maximizing R 2 , which can consider both the prediction error and the explanatory power of the model. 2 The magnitudes of may be different, and directly adding them together may result in one metric dominating the optimization process. Therefore, it may be necessary to normalize these metrics to ensure that each metric contributes to the optimization process in a balanced manner.

[0153] After training and evaluation, an LSTM model that can be used for prediction was obtained. Its input is: the number of preprocessed shield segments, power consumption, and corresponding formation parameters from n moments before the prediction moment to the moment before the prediction moment; the output is: the target carbon emissions of the shield at the prediction moment.

[0154] In machine learning and data preprocessing, normalization is a common technique used to scale all eigenvalues ​​to a uniform scale, usually [0, 1] or [-1, 1]. This helps to increase the convergence speed of the algorithm and improve performance. However, for engineers, it is usually necessary to see the original data scale. Therefore, in a preferred embodiment, denormalization is used, and the denormalization formula of minimum-maximum normalization is:

[0155] x=x'×(x max -x min )+x min

[0156] Among them: x is the original data, x min is the minimum value of a certain type of data in the sample; x max is the maximum value of a certain type of data in the sample, and x' is the normalized data.

[0157] The above-mentioned embodiments can improve the accuracy and precision of the quantification of carbon emissions from shield tunneling in highway tunnels and save time for comparing solutions based on carbon emissions in highway tunnels.

[0158] Based on the same inventive concept, in other embodiments of the present invention, a shield tunneling carbon emission prediction system based on LSTM and bitterfish optimization is also provided, including:

[0159] Data module: obtain the number of segment rings, formation parameters, construction power consumption and actual carbon emissions as training sets;

[0160] Initial modeling module: establish an initial LSTM model and constrain the structural parameter range of the initial LSTM model;

[0161] BFO optimization module: using the BFO method to optimize the structural parameters of the LSTM model to obtain an LSTM model with optimal training parameters;

[0162] Training and evaluation module: using the training set to train the LSTM model with the optimal training parameters, and performing effect evaluation to obtain the LSTM model for prediction;

[0163] Application module: Input the number of segment rings, formation parameters, and construction power consumption in a period of time before the predicted time into the LSTM model used for prediction, obtain the corresponding predicted carbon emission value at the predicted time, and perform inverse normalization processing to obtain the carbon emission value at the original scale.

[0164] The specific implementation techniques of each module / unit in the above example of the present invention can refer to the corresponding steps of the shield tunneling carbon emission prediction method based on LSTM and bitter fish optimization in the above embodiment, which will not be repeated here.

[0165] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various modifications or variations within the scope of the claims, which does not affect the essence of the present invention. The above preferred features can be used in any combination without conflicting with each other.

Claims

1. A shield tunneling carbon emission prediction method based on LSTM and bitterfish optimization, characterized in that: include: The number of segment rings, formation parameters, construction power consumption, and actual carbon emissions are obtained as training sets; Establishing an initial LSTM model and constraining the range of structural parameters of the initial LSTM model; The structural parameters of the LSTM model are optimized using the BFO method to obtain an LSTM model with optimal training parameters; Using the training set, training the LSTM model with the optimal training parameters, and performing effect evaluation to obtain an LSTM model for prediction; The number of segment rings, formation parameters, and construction power consumption in a period of time before the predicted time are input into the prediction LSTM model to obtain the corresponding predicted carbon emission value at the predicted time, and then denormalized to obtain the carbon emission value at the original scale.

2. According to claim 1, a shield tunneling carbon emission prediction method based on LSTM and bitter fish optimization is characterized in that: The number of segment rings, formation parameters, construction power consumption and actual carbon emission values ​​are obtained as a training set, including: Obtain the number of segments in a period of time before the prediction time, as well as the corresponding formation parameters and construction power consumption; Based on the construction power consumption, the actual carbon emission value corresponding to each ring of the segment is obtained by converting the carbon emission coefficient; The segment ring number, formation parameters, construction power consumption and actual carbon emission values ​​are preprocessed as follows: Eliminate the data group corresponding to construction power consumption of 0; Use box plots to process outliers in the selected time period and remove data groups containing outliers; The minimum and maximum normalization methods were used to normalize the eliminated data, and the formation parameters, construction power consumption, and actual carbon emission values ​​corresponding to the number of preprocessed segment rings were obtained.

3. According to claim 1, a shield tunneling carbon emission prediction method based on LSTM and bitter fish optimization is characterized in that: The formation parameters include: cohesion parameters, internal friction angle parameters, penetration resistance parameters, earth pressure parameters and tunnel burial depth parameters of the formation where the segments of the tunnel to be constructed are located.

4. The method for predicting carbon emissions from shield tunneling based on LSTM and bitterfish optimization according to claim 1 is characterized in that: The establishment of an initial LSTM model includes: Add three neural network layers based on the RNN neural network, namely: forget gate, input gate, and output gate, to obtain the initial LSTM model; The initial LSTM model is specifically: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) the t =σ(W o [h t-1 ,x t ]+b o ) h t =o t ·tanh(C t ) Among them, C t-1 is the long-term memory, i.e. the output from the previous unit module, f t is the forgetting factor, h t-1 is short-term memory, W is the weight matrix of the gate, b is the bias term, x t is the event information, the subscript f represents the forget gate, the subscripts i and c represent the first and second parts of the input gate respectively, o represents the output gate, and i t is the long-term memory output from the forget gate, is the short-term memory output from the learning gate, W f is the weight matrix of the forget gate, [h t-1 ,x t ] means connecting two vectors into a longer vector, b f is the bias term of the forget gate, σ is the sigmoid function, W f is the weight matrix of the forget gate, W i is the weight matrix of the input gate, b i is the bias term of the input gate.

5. The method for predicting carbon emissions from shield tunneling based on LSTM and bitterfish optimization according to claim 4 is characterized in that: The structural parameters of the initial LSTM model include: The feature dimension input_size of the input sequence, which specifies the size of the last dimension of the input tensor; The size of the hidden state hidden_size, that is, the output dimension of the LSTM unit, which determines the size of the hidden state in the LSTM unit; The number of LSTM layers num_layers, the default is 1; Whether to use the bias term bias, the default is True; Whether the first dimension of the input and output tensors is the batch size batch_first, the default is False; Whether to perform dropout operation. The default value is 0, which means no dropout operation is applied. Whether to use bidirectional LSTMbidirectional, the default is False.

6. The method for predicting carbon emissions from shield tunneling based on LSTM and bitterfish optimization according to claim 5 is characterized in that: The structural parameter range of the initial LSTM model is constrained, including: The feature dimension input_size of the input sequence is set in the range of 8-12; The size of the hidden state hidden_size is set in the range of 16-64; The number of LSTM layers num_layers is set in the range of 1-6; Whether to use the bias term bias is True; Whether to discard the operation dropout is set to 0.

3.

7. The method for predicting carbon emissions from shield tunneling based on LSTM and bitterfish optimization according to claim 1 is characterized in that: The method of using the BFO method to optimize the structural parameters of the LSTM model to obtain an LSTM model with optimal training parameters includes: Iterative solution: The parameters of LSTM are used as the initial solution in the optimization of BFO method, and each solution represents a bitter fish; Evaluating the fitness of each solution using an objective function; Each solution searches for and attempts to occupy the optimal solution during the iteration process, and updates the position by randomly selecting and competing with the current solution; If the current solution fails to successfully occupy the optimal solution, the current solution does not need to randomly search for a new position; If the optimal solution is occupied, the current solution is spawned, that is, a new solution is generated around the current solution; Evaluate the fitness of each solution and remove solutions that do not meet a set threshold; Update parameters: During the iteration process, the parameters of the BFO method are updated, including: initial population size ips, dimension Dim, initial step size J, initial probability parameter P, exponent a of the decay rate, initial parameter U, initial radius R at spawning, and maximum number of iterations it max , number of test repetitions rp; Iteration termination: When the maximum number of iterations is reached, the iteration stops and the currently found optimal solution is output.

8. The method for predicting carbon emissions from shield tunneling based on LSTM and bitterfish optimization according to claim 7 is characterized in that: The effect of the trained LSTM model is evaluated, including: Calculate the mean square error MSE and sample regression value R between the model prediction value and the true value 2 , as an evaluation indicator of model prediction effect: Where: N is the number of samples, y i is the predicted value of shield target carbon emissions, is the actual value of the shield tunneling target carbon emissions, is the average value of the actual value of the shield target carbon emissions; the smaller the MSE, the higher the R 2 The larger it is, the higher the fitting accuracy of the LSTM prediction model is and the better the prediction performance is.

9. The method for predicting carbon emissions from shield tunneling based on LSTM and bitterfish optimization according to claim 1 is characterized in that: The carbon emission prediction result is denormalized to obtain the carbon emission in the original scale, including: x=x’×(x max -x min )+x min Among them: x is the original data, x min is the minimum value of a type of data in the sample; x max is the maximum value of the same type of data in the sample, and x' is the normalized data.

10. A shield tunneling carbon emission prediction system based on LSTM and bitterfish optimization, characterized in that: include: Data module: obtain the number of segment rings, formation parameters, construction power consumption and actual carbon emissions as training sets; Initial modeling module: establish an initial LSTM model and constrain the structural parameter range of the initial LSTM model; BFO optimization module: using the BFO method to optimize the structural parameters of the LSTM model to obtain an LSTM model with optimal training parameters; Training and evaluation module: using the training set to train the LSTM model with the optimal training parameters, and performing effect evaluation to obtain the LSTM model for prediction; Application module: Input the number of segment rings, formation parameters, and construction power consumption in a period of time before the predicted time into the LSTM model used for prediction, obtain the corresponding predicted carbon emission value at the predicted time, and perform inverse normalization processing to obtain the carbon emission value at the original scale.

Images