A deep learning-based multi-scale rolling prediction method for carbon emissions in the transportation industry

Abstract

The present application relates to the field of energy and environmental information technology, and more particularly to a traffic industry carbon emission multi-scale rolling prediction method based on deep learning, comprising: step 1, data preprocessing and sequence decomposition; step 2, rolling sample construction; step 3, CAWOA initialization; step 4, CAWOA iterative optimization, loop judgment process; step 5, component level BiLSTM training and rolling prediction; step 6, rolling judgment and information update; step 7, mode cycle and prediction output; step 8, mode reconstruction and result restoration. The present application can realize dynamic description and accurate prediction of the evolution trend of traffic industry carbon emission, and provide more realistic prediction basis for relevant emission reduction policy making.

Description

Technical Field

[0001] This invention relates to the field of energy and environmental information technology, and in particular to a multi-scale rolling prediction method for carbon emissions in the transportation industry based on deep learning. Background Technology

[0002] The transportation sector is a significant source of energy consumption and carbon emissions. Its carbon emission levels are influenced by a variety of factors, including energy utilization patterns, transportation scale, economic development, and population structure, exhibiting characteristics such as long-term evolution, strong nonlinearity, and complex fluctuation scales. Accurately predicting the changing trends of carbon emissions in the transportation sector is crucial for achieving energy conservation and emission reduction goals and promoting the green transformation of the transportation industry.

[0003] Currently, the following types of technical methods are mainly used for carbon emission forecasting in the transportation industry:

[0004] 1. Statistical analysis-based prediction methods, such as regression analysis and time series models. These methods are simple in structure and easy to implement, but they usually assume that carbon emission changes have linear characteristics, making it difficult to accurately characterize the nonlinear and abrupt characteristics that are prevalent in actual carbon emission data, resulting in limited prediction accuracy.

[0005] 2. Prediction methods based on traditional machine learning, such as support vector regression and shallow neural networks. These methods improve prediction capabilities to some extent, but they rely on manually setting features and parameters, and the models are insufficient in characterizing long-term dependencies in time series.

[0006] 3. Deep learning-based prediction methods, such as Long Short-Term Memory (LSTM) networks. These methods can automatically learn nonlinear relationships in time series, but they still have the following problems in practical applications: they are sensitive to the non-stationarity of the original carbon emission series, which can easily lead to unstable model training; model performance is highly dependent on hyperparameter configuration, and manual parameter tuning is inefficient and highly subjective; most studies use fixed training sets for prediction and evaluation, without fully considering temporal causal constraints, resulting in deviations between prediction results and real-world application scenarios.

[0007] In addition, most existing technologies directly model the original carbon emission sequence, failing to effectively distinguish the characteristics of change at different time scales, making it difficult for the model to simultaneously take into account both long-term trends and short-term fluctuations.

[0008] To address the aforementioned issues, this application proposes a deep learning-based method for multi-scale rolling prediction of carbon emissions in the transportation industry. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of existing technologies by proposing a multi-scale rolling prediction method for carbon emissions in the transportation industry based on deep learning. This method can dynamically characterize and accurately predict the evolution trend of carbon emissions in the transportation industry, providing a more realistic predictive basis for the formulation of relevant emission reduction policies.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A deep learning-based multi-scale rolling prediction method for carbon emissions in the transportation industry includes the following steps:

[0012] Step 1: Data Preprocessing and Sequence Decomposition

[0013] The annual carbon emissions of the transportation industry are calculated using a carbon emission accounting method based on energy consumption, resulting in a carbon emission time series. After normalizing the original series, VMD is used to decompose it into multiple scales, and the consistency of the decomposition results is checked.

[0014] Step 2: Constructing rolling samples

[0015] Set the lookback window length. When the scroll step is greater than the lookback window length, select a modality to construct an input subsequence and form a training sample of "historical window - next period output".

[0016] Step 3: CAWOA Initialization

[0017] For this mode, the key hyperparameters of BiLSTM are encoded numerically, the population is initialized, the maximum number of iterations is set, and the current optimal solution and its corresponding fitness value are initialized.

[0018] Step 4: CAWOA iterative optimization and iterative judgment process.

[0019] In each iteration, the following steps are performed:

[0020] A) Based on the current individual, construct the corresponding BiLSTM model and calculate the prediction output under the rolling prediction framework;

[0021] B) Calculate the fitness value of an individual based on the rolling prediction error;

[0022] C) Compare the current individual fitness with the historical best value. If the current individual fitness is better than the current best fitness, then update the optimal solution.

[0023] D) Based on the adaptive weights and chaotic perturbation mechanism of CAWOA, update the individual positions of the population and generate a new generation of candidate solutions;

[0024] E) Determine if the termination condition is met: If the maximum number of iterations has not been reached and the convergence condition has not been met, proceed to the next iteration and return to step A); if the maximum number of iterations has been reached or the convergence condition is met, end the CAWOA optimization process and output the optimal hyperparameters of the current mode.

[0025] Step 5: Component-level BiLSTM Training and Rolling Prediction

[0026] Under optimal hyperparameters, retrain the BiLSTM prediction model for this mode; during the rolling step... Output the one-step prediction result and incorporate the actual observations into the sample set;

[0027] Step 6: Scrolling Judgment and Information Update

[0028] Determine if the end of the time series has been reached. If not, scroll the lookback window forward one step, construct a new input sequence, and return to step 5. If the end has been reached, end the rolling prediction process for this mode and obtain its complete prediction sequence.

[0029] Step 7, Modal Cycle and Prediction Output

[0030] Repeat steps 2-6 for all modes to obtain the rolling prediction results for each mode within the entire prediction interval;

[0031] Step 8: Modal reconstruction and result restoration

[0032] At the same time point, the prediction results of each mode are linearly superimposed to obtain the carbon emission prediction value. The prediction sequence is then inversely normalized to obtain the final carbon emission prediction result.

[0033] Preferably, in step 1, a carbon emission accounting method based on energy consumption is used to calculate the annual carbon emissions of the transportation industry:

[0034]

[0035] in, For the first Total annual carbon emissions from the transportation sector; The number of energy types included in the calculation; The energy sources are crude oil, coal, gasoline, fuel oil, natural gas, kerosene, coke, diesel, and electricity. For the transportation industry in the first The annual consumption of the first Energy consumption; For the first The lower heating value of energy sources; Carbon content per unit of calorific value; The oxidation rate during combustion; This is the molecular weight conversion factor for the conversion of carbon to carbon dioxide;

[0036] Based on equation (1), the annual carbon emission time series of the transportation industry can be constructed and used as the target output variable of the subsequent prediction model:

[0037]

[0038] in, This represents the total number of years in the sample.

[0039] Preferably, in step 1, in order to reduce the non-stationarity of the carbon emission series in the transportation industry and extract multi-scale features, variational mode decomposition (VMD) is used to decompose the carbon emission series into multiple scales.

[0040] Let the original carbon emission sequence of the transportation industry be... VMD decomposes it into Modal components And satisfy the refactoring constraints:

[0041]

[0042] in, This is the original carbon emission sequence; For continuous-time variables (corresponding to discrete sampling points in numerical implementation); The number of modes (a positive integer); For the first One modal component;

[0043] To characterize the bandwidth properties of the modes, an analytical signal is first constructed for each mode:

[0044]

[0045] In the formula, For modality Analytical signal (complex value); The Dirac impulse function; The imaginary unit; This is a convolution operation;

[0046] Subsequently, the analytical signal is shifted to a frequency-shifted position. The baseband centered on:

[0047]

[0048] In the formula, The baseband signal after frequency shift (complex value); For the first The central angular frequency of each mode;

[0049] VMD uses the L2 norm of the baseband signal derivative to measure modal bandwidth and minimizes the sum of the modal bandwidths under reconstruction constraints:

[0050]

[0051] In the formula, For the first First mode, To Differentiation operation, It is a norm 2. It is a time series signal;

[0052] To solve this constrained problem, a quadratic penalty coefficient is introduced. With Lagrange multipliers Construct the augmented Lagrangian function:

[0053]

[0054] In the formula, Used to balance bandwidth constraints and reconstruction errors Used to apply reconstruction constraints. Indicates inner product operation;

[0055] Using the alternating direction multiplier method Perform iterative solutions, the first... In the next iteration For iterative indexing, the variable update form is:

[0056]

[0057] In the formula, For the first The modal component in the th ... Update value at the next iteration For the first The mode in the th ... The center frequency at the next iteration For modality The frequency domain representation (Fourier transform). For frequency variables, For the first The Lagrange multipliers of the next iteration, Update the step size for the multiplier.

[0058] The algorithm terminates when the convergence criterion is met:

[0059]

[0060] In the formula, This is the preset convergence threshold.

[0061] Preferably, in steps 3 and 4, the improved whale optimization algorithm (CAWOA) is introduced to achieve global optimization of the hyperparameters of the BiLSTM prediction model.

[0062] Let the dimension of the parameter space to be optimized be . Population size is , No. The era The position vector of an individual is represented as:

[0063]

[0064] in, Indicates the first The hyperparameter at the th ... The value of the substitute, , Let the maximum number of iterations be denoted by and the global optimal solution of the current population be denoted by . ;

[0065] To dynamically adjust the algorithm's exploration and development capabilities during the search process, a nonlinear adaptive weight function is introduced:

[0066]

[0067] in, As the iteration number changes, it is used to control the degree to which an individual's update depends on the current optimal solution;

[0068] In the position update process, a control vector is introduced. and Its definition is:

[0069]

[0070] in, For a random vector, the convergence factor is... Decreases linearly with iteration:

[0071]

[0072] When the algorithm is in the local development stage, the individual searches around the current optimal solution, and its position is updated as follows:

[0073]

[0074] in, This represents element-wise multiplication. This indicates taking the absolute value of each element;

[0075] During the spiral search phase, the individual approaches the optimal solution along an exponential spiral trajectory, and its update form is as follows:

[0076]

[0077] in, For spiral shape control parameters, It is a random number;

[0078] when At this point, the algorithm enters the global exploration phase, randomly selecting individuals. Guided search, its location is updated to:

[0079]

[0080] in, Individuals randomly selected from the current population;

[0081] To enhance the algorithm's ability to escape local optima, when the global optimum is in continuous... When no update occurs within a generation, a chaotic perturbation mechanism is introduced. Perturbations are applied, and a chaotic sequence is generated using a Logistic mapping:

[0082]

[0083] For the Decision variables First, normalization is performed:

[0084]

[0085] Obtained through chaotic mapping Then map it back to the original search space:

[0086]

[0087] Through the aforementioned adaptive weight adjustment, improved position update strategy, and chaotic perturbation mechanism, CAWOA is used to globally optimize the hyperparameters of BiLSTM networks in a complex parameter space.

[0088] Preferably, in step 5, BiLSTM deep learning is used to characterize the nonlinear time-series mapping relationship between driving variables and carbon emission modal components. A bidirectional long short-term memory network is used as the prediction model. BiLSTM improves the ability to characterize the dynamic features of time series by simultaneously modeling the positive and negative dependencies of the sequence.

[0089] Let the input time series be ,in This represents the time step index within the network. Given the length of the input sequence, For the first The input vector at each time step is represented in an LSTM unit by the forget gate, input gate, output gate, and candidate memory unit as follows:

[0090]

[0091] in, , , and They are respectively The forgetting gate, input gate, output gate, and unit state of memory cells at any given moment. This represents the Sigmoid activation function. It is the hyperbolic tangent function. and This is the weight matrix. For bias vectors, This is the implicit output of the previous time step;

[0092] Memory state With implicit output The update format is as follows:

[0093]

[0094] in, This represents element-wise multiplication. This represents the memory state at the s-th time step. This represents the implicit output at the s-th time step;

[0095] BiLSTM consists of two information streams, one forward and one backward, with the forward stream implicitly containing the output. With inverse implicit output They are obtained from the following two formulas respectively:

[0096]

[0097] Received, among which and These represent the recursive operators for the forward and reverse LSTM units, respectively.

[0098] By concatenating the bidirectional implicit outputs, we obtain a bidirectional representation vector:

[0099]

[0100] The sequence-level representation is formed by taking the last time step, and the network output is obtained through a fully connected layer.

[0101]

[0102] in, and For the output layer weights and biases, For the first The network prediction output for each sample The sample number is [sample ID], and the total number of training samples is [number]. ;

[0103] The model training uses the mean squared error loss function:

[0104]

[0105] in Let be the true output value corresponding to the p-th sample. To reduce the risk of overfitting, a random deactivation mechanism is introduced into the hidden layer, and its deactivation ratio is used as a hyperparameter to participate in the subsequent optimization process.

[0106] Preferably, in step 6, to ensure that the time causality relationship is strictly followed during the time series forecasting process, a rolling forecasting method is used to model and evaluate the carbon emission series. Let the carbon emission time series be... ,in A discrete index representing a time series (corresponding to the year). This represents the total length of the sample.

[0107] In the scrolling prediction framework, the lookback window length is set to... In the nth prediction, based on the historical observation sequence As input, construct a prediction model and generate predictions for the next time point:

[0108]

[0109] in, Indicates the first The predicted mappings trained in the next rolling step For the first Predicted carbon emission values ​​at specific time points;

[0110] After completing a prediction, the actual observed values ​​will be... Incorporate the sample set and scroll the lookback window forward one time step to form a new input sequence:

[0111]

[0112] The above process is repeated continuously until the end of the sequence, thus achieving a stepwise prediction of the entire time interval.

[0113] By adopting the above technical solution: First, an energy consumption-based carbon emission accounting method is used to calculate the annual carbon emissions of the transportation industry; second, a variational mode decomposition method is used to decompose the carbon emission sequence of the transportation industry into multiple scales, effectively improving the predictability of subsequences; third, an improved whale optimization algorithm is introduced to globally optimize the key hyperparameters of the bidirectional long short-term memory network, enhancing the model's predictive stability under small sample conditions; finally, under strict adherence to time causality constraints, a fixed window rolling prediction mechanism is introduced to recursively predict each modal component and reconstruct the overall carbon emission sequence. This method can achieve dynamic characterization and accurate prediction of the evolution trend of carbon emissions in the transportation industry, providing a more realistic predictive basis for the formulation of relevant emission reduction policies.

[0114] Compared with the prior art, the present invention has the following beneficial effects:

[0115] 1. To achieve effective integration of energy consumption characteristics and deep learning models, this invention introduces energy consumption-related indicators of the transportation industry into the model input layer, enabling the deep learning model to simultaneously learn the intrinsic relationship between the temporal evolution characteristics of carbon emissions and energy utilization characteristics. This avoids the problem of insufficient information caused by relying solely on a single carbon emission sequence for prediction, and improves the rationality and interpretability of the prediction results.

[0116] 2. By reducing the non-stationarity of carbon emission sequences through multi-scale decomposition, this invention uses a multi-scale decomposition method to preprocess carbon emission sequences in the transportation industry, decomposing the original complex sequence into several relatively stationary subsequences, enabling deep learning models to model at different time scales, thereby effectively mitigating the negative impact of non-stationarity on prediction accuracy and model stability.

[0117] 3. Enhance the ability of deep learning models to comprehensively depict long-term and short-term changes. Through multi-scale modeling and result reconstruction mechanisms, this invention can simultaneously depict the long-term trend and short-term fluctuation characteristics of carbon emissions in the transportation industry, avoiding the problem that traditional single-scale models can only focus on local changes or overall trends.

[0118] 4. Improve the adaptability and predictive stability of model parameter configuration. This invention introduces an intelligent optimization algorithm to automatically optimize the key parameters of the deep learning model, reducing the subjectivity and repetitiveness of manual parameter tuning, and improving the predictive stability and robustness of the model under different data stages.

[0119] 5. Enhance the practical application value of prediction results through a rolling prediction mechanism. This invention adopts a fixed window rolling prediction method, which dynamically updates the training sample set at each prediction time, strictly follows the temporal causal relationship, avoids the problem of future information leakage, and makes the prediction process and prediction results closer to the actual application scenario. It is suitable for continuously updated carbon emission prediction tasks in the transportation industry.

[0120] 6. Applicable to medium- and long-term carbon emission trend analysis and dynamic assessment. This invention can predict carbon emissions in the transportation industry period by period based on historical data. It is suitable for judging medium- and long-term carbon emission trends and can provide technical support for the assessment of carbon emission reduction targets, energy structure adjustment and policy formulation in the transportation industry.

[0121] 7. The method is highly versatile and easy to implement and promote in engineering. This invention does not depend on a specific region or a single data source. The multi-scale decomposition, deep learning modeling and rolling prediction steps can be flexibly configured according to the actual data situation. It has good versatility and scalability, and is easy to promote and apply in different transportation systems or energy management scenarios. Attached Figure Description

[0122] Figure 1 This is an overall flowchart of the present invention;

[0123] Figure 2 This is a flowchart of the CAWOA algorithm in this invention;

[0124] Figure 3 This is a diagram of the BiLSTM deep learning structure in this invention;

[0125] Figure 4 This is a schematic diagram of the rolling mechanism in this invention;

[0126] Figure 5 This is a comparison chart of the multi-model rolling prediction results of an embodiment of the present invention. Detailed Implementation

[0127] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings, so that those skilled in the art can better understand the advantages and features of the present invention, thereby making a clearer definition of the scope of protection of the present invention. The embodiments described in this invention are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0128] Example:

[0129] Data on carbon emissions and their main influencing factors in the transportation industry from 1990 to 2019 were selected as the research sample. Relevant data were obtained from the National Statistical Yearbook and authoritative statistical sources. To eliminate differences in the units of measurement of different variables and improve model stability, all input and output variables were normalized. Simultaneously, data integrity and consistency were checked to ensure that no missing or outlier values ​​interfered with model training and prediction results.

[0130] Based on the formation mechanism of carbon emissions in the transportation industry and related research findings, an input variable system was constructed from three dimensions. The definitions and statistical scope of each variable are shown in Table 1. This variable system can comprehensively reflect the driving characteristics of carbon emissions in the transportation industry, providing a basic support for the construction of prediction models.

[0131] Table 1. Variable Definitions and Statistical Definitions

[0132]

[0133] A deep learning-based multi-scale rolling prediction method for carbon emissions in the transportation industry includes the following steps:

[0134] Step 1: Data Preprocessing and Sequence Decomposition

[0135] The annual carbon emissions of the transportation industry are calculated using a carbon emission accounting method based on energy consumption, resulting in a carbon emission time series. After normalizing the original series, VMD is used to decompose it into multiple scales, and the consistency of the decomposition results is checked.

[0136] Step 2: Constructing rolling samples

[0137] Set the lookback window length. When the scroll step is greater than the lookback window length, select a modality to construct an input subsequence and form a training sample of "historical window - next period output".

[0138] Step 3: CAWOA Initialization

[0139] For this mode, the key hyperparameters of BiLSTM are encoded numerically, the population is initialized, the maximum number of iterations is set, and the current optimal solution and its corresponding fitness value are initialized.

[0140] Step 4: CAWOA iterative optimization and iterative judgment process.

[0141] In each iteration, the following steps are performed:

[0142] A) Based on the current individual, construct the corresponding BiLSTM model and calculate the prediction output under the rolling prediction framework;

[0143] B) Calculate the fitness value of an individual based on the rolling prediction error;

[0144] C) Compare the current individual fitness with the historical best value. If the current individual fitness is better than the current best fitness, then update the optimal solution.

[0145] D) Based on the adaptive weights and chaotic perturbation mechanism of CAWOA, update the individual positions of the population and generate a new generation of candidate solutions;

[0146] E) Determine if the termination condition is met: If the maximum number of iterations has not been reached and the convergence condition has not been met, proceed to the next iteration and return to step A); if the maximum number of iterations has been reached or the convergence condition is met, end the CAWOA optimization process and output the optimal hyperparameters of the current mode.

[0147] Step 5: Component-level BiLSTM Training and Rolling Prediction

[0148] Under optimal hyperparameters, retrain the BiLSTM prediction model for this mode; during the rolling step... Output the one-step prediction result and incorporate the actual observations into the sample set;

[0149] Step 6: Scrolling Judgment and Information Update

[0150] Determine if the end of the time series has been reached. If not, scroll the lookback window forward one step, construct a new input sequence, and return to step 5. If the end has been reached, end the rolling prediction process for this mode and obtain its complete prediction sequence.

[0151] Step 7, Modal Cycle and Prediction Output

[0152] Repeat steps 2-6 for all modes to obtain the rolling prediction results for each mode within the entire prediction interval;

[0153] Step 8: Modal reconstruction and result restoration

[0154] At the same time point, the prediction results of each mode are linearly superimposed to obtain the carbon emission prediction value. The prediction sequence is then inversely normalized to obtain the final carbon emission prediction result.

[0155] Specifically, in step 1, carbon emissions from the transportation sector mainly originate from the combustion of fossil fuels during transportation activities, and their scale is directly affected by the energy consumption structure and intensity. Therefore, a carbon emission accounting method based on energy consumption is used to calculate the annual carbon emissions of the transportation sector:

[0156]

[0157] in, For the first Total annual carbon emissions from the transportation sector; The number of energy types included in the calculation; The energy sources are crude oil, coal, gasoline, fuel oil, natural gas, kerosene, coke, diesel, and electricity. For the transportation industry in the first The annual consumption of the first Energy consumption; For the first The lower heating value of energy sources; Carbon content per unit of calorific value; The oxidation rate during combustion; This is the molecular weight conversion factor for the conversion of carbon to carbon dioxide;

[0158] Based on equation (1), the annual carbon emission time series of the transportation industry can be constructed and used as the target output variable of the subsequent prediction model:

[0159]

[0160] in, This represents the total number of years in the sample.

[0161] Specifically, in step 1, to reduce the non-stationarity of the carbon emission sequence in the transportation industry and extract multi-scale features, Variational Mode Decomposition (VMD) is used to decompose the carbon emission sequence into multiple scales. VMD aims to decompose the original sequence into several modal components with finite bandwidth and simultaneously estimate the center frequency corresponding to each mode, so that the decomposition results can reconstruct the original sequence without significant loss of information.

[0162] Let the original carbon emission sequence of the transportation industry be... VMD decomposes it into Modal components And satisfy the refactoring constraints:

[0163]

[0164] in, This is the original carbon emission sequence; For continuous-time variables (corresponding to discrete sampling points in numerical implementation); The number of modes (a positive integer); For the first One modal component;

[0165] To characterize the bandwidth properties of the modes, an analytical signal is first constructed for each mode:

[0166]

[0167] In the formula, For modality Analytical signal (complex value); The Dirac impulse function; The imaginary unit; This is a convolution operation;

[0168] Subsequently, the analytical signal is shifted to a frequency-shifted position. The baseband centered on:

[0169]

[0170] In the formula, The baseband signal after frequency shift (complex value); For the first The central angular frequency of each mode;

[0171] VMD uses the L2 norm of the baseband signal derivative to measure modal bandwidth and minimizes the sum of the modal bandwidths under reconstruction constraints:

[0172]

[0173] In the formula, For the first First mode, To Differentiation operation, It is a norm 2. It is a time series signal;

[0174] To solve this constrained problem, a quadratic penalty coefficient is introduced. With Lagrange multipliers Construct the augmented Lagrangian function:

[0175]

[0176] In the formula, Used to balance bandwidth constraints and reconstruction errors Used to apply reconstruction constraints. Indicates inner product operation;

[0177] Using the alternating direction multiplier method Perform iterative solutions, the first... In the next iteration For iterative indexing, the variable update form is:

[0178]

[0179] In the formula, For the first The modal component in the th ... Update value at the next iteration For the first The mode in the th ... The center frequency at the next iteration For modality The frequency domain representation (Fourier transform). For frequency variables, For the first The Lagrange multipliers of the next iteration, Update the step size for the multiplier.

[0180] The algorithm terminates when the convergence criterion is met:

[0181]

[0182] In the formula, This is the preset convergence threshold.

[0183] Specifically, in steps 3 and 4, the improved whale optimization algorithm (CAWOA) is introduced to achieve global optimization of the hyperparameters of the BiLSTM prediction model.

[0184] Let the dimension of the parameter space to be optimized be . Population size is , No. The era The position vector of an individual is represented as:

[0185]

[0186] in, Indicates the first The hyperparameter at the th ... The value of the substitute, , Let the maximum number of iterations be denoted by and the global optimal solution of the current population be denoted by . ;

[0187] To dynamically adjust the algorithm's exploration and development capabilities during the search process, a nonlinear adaptive weight function is introduced:

[0188]

[0189] in, As the iteration number changes, it is used to control the degree to which an individual's update depends on the current optimal solution;

[0190] In the position update process, a control vector is introduced. and Its definition is:

[0191]

[0192] in, For a random vector, the convergence factor is... Decreases linearly with iteration:

[0193]

[0194] When the algorithm is in the local development stage, the individual searches around the current optimal solution, and its position is updated as follows:

[0195]

[0196] in, This represents element-wise multiplication. This indicates taking the absolute value of each element;

[0197] During the spiral search phase, the individual approaches the optimal solution along an exponential spiral trajectory, and its update form is as follows:

[0198]

[0199] in, For spiral shape control parameters, It is a random number;

[0200] when At this point, the algorithm enters the global exploration phase, randomly selecting individuals. Guided search, its location is updated to:

[0201]

[0202] in, Individuals randomly selected from the current population;

[0203] To enhance the algorithm's ability to escape local optima, when the global optimum is in continuous... When no update occurs within a generation, a chaotic perturbation mechanism is introduced. Perturbations are applied, and a chaotic sequence is generated using a Logistic mapping:

[0204]

[0205] For the Decision variables First, normalization is performed:

[0206]

[0207] Obtained through chaotic mapping Then map it back to the original search space:

[0208]

[0209] Through the aforementioned adaptive weight adjustment, improved position update strategy, and chaotic perturbation mechanism, CAWOA is used to globally optimize the hyperparameters of BiLSTM networks in a complex parameter space.

[0210] Specifically, in step 5, BiLSTM deep learning is used to characterize the nonlinear time-series mapping relationship between driving variables and carbon emission modal components. A bidirectional long short-term memory network is used as the prediction model. By simultaneously modeling the positive and negative dependencies of the sequence, BiLSTM improves the ability to characterize the dynamic features of the time series.

[0211] Let the input time series be ,in This represents the time step index within the network. Given the length of the input sequence, For the first The input vector at each time step is represented in an LSTM unit by the forget gate, input gate, output gate, and candidate memory unit as follows:

[0212]

[0213] in, , , and They are respectively The forgetting gate, input gate, output gate, and unit state of memory cells at any given moment. This represents the Sigmoid activation function. It is the hyperbolic tangent function. and This is the weight matrix. For bias vectors, This is the implicit output of the previous time step;

[0214] Memory state With implicit output The update format is as follows:

[0215]

[0216] in, This represents element-wise multiplication. This represents the memory state at the s-th time step. This represents the implicit output at the s-th time step;

[0217] BiLSTM consists of two information streams, one forward and one backward, with the forward stream implicitly containing the output. With inverse implicit output They are obtained from the following two formulas respectively:

[0218]

[0219] Received, among which and These represent the recursive operators for the forward and reverse LSTM units, respectively.

[0220] By concatenating the bidirectional implicit outputs, we obtain a bidirectional representation vector:

[0221]

[0222] The sequence-level representation is formed by taking the last time step, and the network output is obtained through a fully connected layer.

[0223]

[0224] in, and For the output layer weights and biases, For the first The network prediction output for each sample The sample number is [sample ID], and the total number of training samples is [number]. ;

[0225] The model training uses the mean squared error loss function:

[0226]

[0227] in Let be the true output value corresponding to the p-th sample. To reduce the risk of overfitting, a random deactivation mechanism is introduced into the hidden layer, and its deactivation ratio is used as a hyperparameter to participate in the subsequent optimization process.

[0228] Specifically, in step 6, to ensure that the time series forecasting process strictly adheres to the temporal causal relationship, a rolling forecasting method is used to model and evaluate the carbon emission series. Let the carbon emission time series be... ,in A discrete index representing a time series (corresponding to the year). This represents the total length of the sample.

[0229] In the scrolling prediction framework, the lookback window length is set to... In the nth prediction, based on the historical observation sequence As input, construct a prediction model and generate predictions for the next time point:

[0230]

[0231] in, Indicates the first The predicted mappings trained in the next rolling step For the first Predicted carbon emission values ​​at specific time points;

[0232] After completing a prediction, the actual observed values ​​will be... Incorporate the sample set and scroll the lookback window forward one time step to form a new input sequence:

[0233]

[0234] The above process is repeated continuously until the end of the sequence, thus achieving a stepwise prediction of the entire time interval.

[0235] To verify the predictive advantage of the model, BP neural network, support vector regression (SVR), long short-term memory network (LSTM) and WOA-LSTM model were selected as comparison methods and a multi-index joint evaluation was conducted. The results are shown in Table 2.

[0236] Table 2 Evaluation Results of Multi-Model Rolling Prediction

[0237]

[0238] As shown in Table 2, the model achieved optimal results across all evaluation metrics, with RMSE, MAE, and MAPE values ​​of 10.9468, 6.9431, and 0.0460, respectively. The accuracy reached 0.9699, significantly outperforming the comparison models. Compared to BP, SVR, LSTM, and WOA-LSTM, this model showed significant advantages in error control and fitting accuracy, indicating that it has higher prediction stability and generalization ability under rolling prediction conditions.

[0239] To visually demonstrate the predictive performance of different models, Figure 5 shows a comparison between the predicted results and the actual values. It can be observed that the model exhibits strong consistency in both overall trend fitting and local fluctuation characterization, and is able to track the changing trajectory of carbon emissions in the transportation sector effectively.

[0240] The above results demonstrate that the present invention can effectively mitigate the adverse effects of non-stationary time series on prediction models, improve the stability and reliability of rolling prediction results, and has good application potential in complex time series prediction and carbon emission analysis in the transportation industry. It can also provide quantitative support for relevant policy evaluation and scenario analysis.

[0241] The descriptions and practices disclosed in this invention are readily apparent and understandable to those skilled in the art, and various modifications and refinements can be made without departing from the principles of this invention. Therefore, any modifications or improvements made without departing from the spirit of this invention should also be considered within the scope of protection of this invention.

Claims

1. A multi-scale rolling prediction method for carbon emissions in the transportation industry based on deep learning, characterized in that, Includes the following steps: Step 1: Data Preprocessing and Sequence Decomposition The annual carbon emissions of the transportation industry are calculated using a carbon emission accounting method based on energy consumption, resulting in a carbon emission time series. After normalizing the original series, VMD is used to decompose it into multiple scales, and the consistency of the decomposition results is checked. Step 2: Constructing rolling samples Set the lookback window length. When the scroll step is greater than the lookback window length, select a modality to construct an input subsequence and form a training sample of "historical window - next period output". Step 3: CAWOA Initialization For this mode, the key hyperparameters of BiLSTM are encoded numerically, the population is initialized, the maximum number of iterations is set, and the current optimal solution and its corresponding fitness value are initialized. Step 4: CAWOA iterative optimization and iterative judgment process. In each iteration, the following steps are performed: A) Based on the current individual, construct the corresponding BiLSTM model and calculate the prediction output under the rolling prediction framework; B) Calculate the fitness value of an individual based on the rolling prediction error; C) Compare the current individual fitness with the historical best value. If the current individual fitness is better than the current best fitness, then update the optimal solution. D) Based on the adaptive weights and chaotic perturbation mechanism of CAWOA, update the individual positions of the population and generate a new generation of candidate solutions; E) Determine if the termination condition is met: If the maximum number of iterations has not been reached and the convergence condition has not been met, proceed to the next iteration and return to step A); If the maximum number of iterations is reached or the convergence condition is met, the CAWOA optimization process ends, and the optimal hyperparameters of the current mode are output. Step 5: Component-level BiLSTM Training and Rolling Prediction Under optimal hyperparameters, retrain the BiLSTM prediction model for this mode; during the rolling step... Output the one-step prediction result and incorporate the actual observations into the sample set; Step 6: Scrolling Judgment and Information Update Determine if the end of the time series has been reached. If not, scroll the lookback window forward one step, construct a new input sequence, and return to step 5. If the end has been reached, end the rolling prediction process for this mode and obtain its complete prediction sequence. Step 7, Modal Cycle and Prediction Output Repeat steps 2-6 for all modes to obtain the rolling prediction results for each mode within the entire prediction interval; Step 8: Modal reconstruction and result restoration At the same time point, the prediction results of each mode are linearly superimposed to obtain the carbon emission prediction value. The prediction sequence is then inversely normalized to obtain the final carbon emission prediction result.

2. The method for multi-scale rolling prediction of carbon emissions in the transportation industry based on deep learning according to claim 1, characterized in that, In step 1, a carbon emission accounting method based on energy consumption is used to calculate the annual carbon emissions of the transportation industry: in, For the first Total annual carbon emissions from the transportation sector; The number of energy types included in the calculation; The energy sources are crude oil, coal, gasoline, fuel oil, natural gas, kerosene, coke, diesel, and electricity. For the transportation industry in the first The annual consumption of the first Energy consumption; For the first The lower heating value of energy sources; Carbon content per unit of calorific value; The oxidation rate during combustion; This is the molecular weight conversion factor for the conversion of carbon to carbon dioxide; Based on equation (1), the annual carbon emission time series of the transportation industry is constructed and used as the target output variable of the subsequent prediction model: in, This represents the total number of years in the sample.

3. The method for multi-scale rolling prediction of carbon emissions in the transportation industry based on deep learning according to claim 1, characterized in that, In step 1, in order to reduce the non-stationarity of the carbon emission series in the transportation industry and extract multi-scale features, variational mode decomposition (VMD) is used to decompose the carbon emission series into multiple scales. Let the original carbon emission sequence of the transportation industry be... VMD decomposes it into Modal components And satisfy the refactoring constraints: in, This is the original carbon emission sequence; For continuous time variables; The number of modes; For the first One modal component; To characterize the bandwidth properties of the modes, an analytical signal is first constructed for each mode: In the formula, For modality Analyzed signals; The Dirac impulse function; The imaginary unit; This is a convolution operation; Subsequently, the analytical signal is shifted to a frequency-shifted position. The baseband centered on: In the formula, The baseband signal after frequency shifting; For the first The central angular frequency of each mode; VMD uses the L2 norm of the baseband signal derivative to measure modal bandwidth and minimizes the sum of the modal bandwidths under reconstruction constraints: In the formula, For the first First mode, To Differentiation operation, It is a norm 2. It is a time series signal; To solve this constrained problem, a quadratic penalty coefficient is introduced. With Lagrange multipliers Construct the augmented Lagrangian function: In the formula, Used to balance bandwidth constraints and reconstruction errors Used to apply reconstruction constraints. Indicates inner product operation; Using the alternating direction multiplier method Perform iterative solutions, the first... In the next iteration For iterative indexing, the variable update form is: In the formula, For the first The modal component in the th ... Update value at the next iteration For the first The mode in the th ... The center frequency at the next iteration For modality The frequency domain representation, For frequency variables, For the first The Lagrange multipliers of the next iteration, Update the step size for the multiplier; The algorithm terminates when the convergence criterion is met: In the formula, This is the preset convergence threshold.

4. The multi-scale rolling prediction method for carbon emissions in the transportation industry based on deep learning according to claim 1, characterized in that, In steps 3 and 4, the improved whale optimization algorithm CAWOA is introduced to achieve global optimization of the hyperparameters of the BiLSTM prediction model. Let the dimension of the parameter space to be optimized be . Population size is , No. The era The position vector of an individual is represented as: in, Indicates the first The hyperparameter at the th ... The value of the substitute, , Let the maximum number of iterations be denoted by and the global optimal solution of the current population be denoted by . ; To dynamically adjust the algorithm's exploration and development capabilities during the search process, a nonlinear adaptive weight function is introduced: in, As the iteration number changes, it is used to control the degree to which an individual's update depends on the current optimal solution; In the position update process, a control vector is introduced. and Its definition is: in, For a random vector, the convergence factor is... Decreases linearly with iteration: When the algorithm is in the local development stage, the individual searches around the current optimal solution, and its position is updated as follows: in, This represents element-wise multiplication. This indicates taking the absolute value of each element; During the spiral search phase, the individual approaches the optimal solution along an exponential spiral trajectory, and its update form is as follows: in, For spiral shape control parameters, It is a random number; when At this point, the algorithm enters the global exploration phase, randomly selecting individuals. Guided search, its location is updated to: in, Individuals randomly selected from the current population; To enhance the algorithm's ability to escape local optima, when the global optimum is in continuous... When no update occurs within a generation, a chaotic perturbation mechanism is introduced. Perturbations are applied, and a chaotic sequence is generated using a Logistic mapping: For the Decision variables First, normalization is performed: Obtained through chaotic mapping Then map it back to the original search space: Through the aforementioned adaptive weight adjustment, improved position update strategy, and chaotic perturbation mechanism, CAWOA is used to globally optimize the hyperparameters of BiLSTM networks in a complex parameter space.

5. The multi-scale rolling prediction method for carbon emissions in the transportation industry based on deep learning according to claim 1, characterized in that, In step 5, BiLSTM deep learning is used to characterize the nonlinear time-series mapping relationship between driving variables and carbon emission modal components. A bidirectional long short-term memory network is used as the prediction model. By simultaneously modeling the positive and negative dependencies of the sequence, BiLSTM improves the ability to characterize the dynamic features of the time series. Let the input time series be ,in This represents the time step index within the network. Given the length of the input sequence, For the first The input vector at each time step is represented in an LSTM unit by the forget gate, input gate, output gate, and candidate memory unit as follows: in, , , and They are respectively The forgetting gate, input gate, output gate, and unit state of memory cells at any given moment. This represents the Sigmoid activation function. It is the hyperbolic tangent function. and This is the weight matrix. For bias vectors, This is the implicit output of the previous time step; Memory state With implicit output The update format is as follows: in, This represents element-wise multiplication. This represents the memory state at the s-th time step. This represents the implicit output at the s-th time step; BiLSTM consists of two information streams, one forward and one backward, with the forward stream implicitly containing the output. With inverse implicit output They are obtained from the following two formulas respectively: Received, among which and These represent the recursive operators for the forward and reverse LSTM units, respectively. By concatenating the bidirectional implicit outputs, we obtain a bidirectional representation vector: The sequence-level representation is formed by taking the "last time step", and the network output is obtained through a fully connected layer: in, and For the output layer weights and biases, For the first The network prediction output for each sample The sample number is [sample ID], and the total number of training samples is [number]. ; The model training uses the mean squared error loss function: in Let be the true output value corresponding to the p-th sample. To reduce the risk of overfitting, a random deactivation mechanism is introduced into the hidden layer, and its deactivation ratio is used as a hyperparameter to participate in the subsequent optimization process.

6. The multi-scale rolling prediction method for carbon emissions in the transportation industry based on deep learning according to claim 1, characterized in that, In step 6, to ensure that the time series forecasting process strictly adheres to the temporal causal relationship, a rolling forecasting method is used to model and evaluate the carbon emission series. Let the carbon emission time series be... ,in Discrete indexes representing time series. This represents the total length of the sample. In the scrolling prediction framework, the lookback window length is set to... In the nth prediction, based on the historical observation sequence As input, construct a prediction model and generate predictions for the next time point: in, Indicates the first The predicted mappings trained in the next rolling step For the first Predicted carbon emission values ​​at specific time points; After completing a prediction, the actual observed values ​​will be... Incorporate the sample set and scroll the lookback window forward one time step to form a new input sequence: The above process is repeated continuously until the end of the sequence, thus achieving a stepwise prediction of the entire time interval.

Images