Combined carbon price prediction method based on decomposition integration and BiLSTM model

Abstract

The invention discloses a combined carbon price prediction method based on decomposition integration and a BiLSTM model. The method comprises the following steps: firstly, collecting carbon price time sequence data of a carbon emission right center and carrying out data cleaning; then decomposing the cleaned carbon valence time sequence data into a high-frequency sub-sequence and a low-frequency sub-sequence by adopting a variational mode decomposition method; meanwhile, analyzing correlation coefficients and average periods of the high-frequency subsequences and the low-frequency subsequences by utilizing a Pearson correlation analysis method, and verifying a decomposition result of the variational mode decomposition method; and finally, different DLUBE neural network models are selected according to the fluctuation forms of different subsequences to carry out carbon price prediction, and a combined carbon price interval prediction result is obtained. By introducing the decomposition integration strategy, the original time sequence is decomposed into the low-frequency sequence and the high-frequency sequence, and combined prediction is performed according to different fluctuation forms and data features, so that the influence of abnormal value mixed noise is reduced, and the efficiency and accuracy are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning and carbon price prediction, and specifically relates to a combined carbon price prediction method based on decomposition integration and a BiLSTM model. Background Art

[0002] Carbon price forecasting plays a vital role in the carbon trading market. It is helpful to understand the dynamics of the carbon market and promote enterprises to reduce carbon emissions and low-carbon transformation. Common carbon price forecasting methods include econometric models and machine learning methods. With the continuous development and improvement of machine learning theory, deep learning methods such as artificial neural networks, recursive neural networks, and convolutional neural networks have also been applied to carbon price forecasting.

[0003] At present, the research on carbon price forecasting at home and abroad is divided into point forecasting and interval forecasting according to the different forms of output. Different from point forecasting, interval forecasting is to quantify the possible uncertainty of point forecasting. It consists of upper and lower limits, providing the possible range of the target. For decision makers, this is more reliable and more informative than point forecasting results. At present, there are three main methods for interval forecasting: one is to achieve interval forecasting through deterministic error distribution and confidence level on the basis of point forecasting; the second is to use interval value data to predict the upper and lower limits of interval data respectively, and enrich the input information to reduce the randomness of the interval generated by a single sequence; the third is to directly obtain the upper and lower limits of the prediction interval (upper and lower limit estimation method, LUBE). This method is a data-driven non-parametric solution. It does not assume any specific distribution function to directly provide the upper and lower limits of the interval. The purpose is to more realistically describe the uncertainty of deterministic forecasting; but it also has defects: first, its objective function is based on qualitative and has no statistical significance; second, the integrated loss function cannot fully represent the interval quality information measured by two conflicting indicators, resulting in potential performance loss.

[0004] Since the deep learning method has more advantages in predicting time series, it can better learn the complex nonlinear relationship between input variables, and the parameter-free method can also reduce human errors, making the generated prediction interval more intuitive and effective. Therefore, the present invention adopts the deep learning method for interval prediction. Summary of the invention

[0005] The main purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art and provide a combined carbon price prediction method based on decomposition integration and BiLSTM model. By introducing a decomposition integration strategy, the original time series is decomposed into low-frequency series and high-frequency series, and combined prediction is performed according to their different fluctuation forms and data characteristics to reduce the influence of outlier mixed noise. Finally, the carbon price is predicted based on the combined interval prediction model of genetic algorithm to improve efficiency and accuracy.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The combined carbon price prediction method based on decomposition integration and BiLSTM model includes the following steps:

[0008] Collect carbon price time series data from the carbon emission rights center over a period of time;

[0009] Perform data cleaning on carbon price time series data;

[0010] The variational mode decomposition method is used to decompose the cleaned carbon price time series data into high-frequency subsequences and low-frequency subsequences. At the same time, the Pearson correlation analysis method is used to analyze the correlation coefficients and average periods of the high-frequency subsequences and low-frequency subsequences to verify the decomposition results of the variational mode decomposition method.

[0011] Different DLUBE neural network models are selected for carbon price prediction according to the fluctuation forms of different subsequences to obtain a combined carbon price range prediction result; the different DLUBE neural network models include a BP model and a BiLSTM model; the different DLUBE neural network models are all obtained by optimizing parameters using a genetic algorithm.

[0012] As a preferred technical solution, the data cleaning adopts the maximum value normalization method.

[0013] As a preferred technical solution, the variational mode decomposition method is used to decompose the cleaned carbon price time series data into high-frequency subsequences and low-frequency subsequences, specifically:

[0014] The cleaned carbon price time series data is used to calculate the orthogonal components of the signal through Hilport transformation to obtain the analytical signal;

[0015] The objective function of the variational mode decomposition method is constructed based on the analytical signal, and the constrained optimization objective function is constructed using the Lagrange multiplier method;

[0016] The optimized objective function is iteratively solved by the alternating direction multiplier method to obtain several modes and their corresponding center frequencies;

[0017] Several modes are classified according to the center frequency or spectrum to obtain high-frequency subsequences and low-frequency subsequences.

[0018] As a preferred technical solution, the analysis of the correlation coefficient and average period of the high-frequency subsequence and the low-frequency subsequence is specifically as follows:

[0019] The Pearson correlation coefficient is used to analyze the correlation between the high-frequency subsequence and the low-frequency subsequence and the cleaned carbon price time series data, and the correlation coefficient is calculated. If the correlation coefficient is lower than the set coefficient threshold, the cleaned carbon price time series data is decomposed again;

[0020] Use the peak detection algorithm to find the peak position of each high-frequency subsequence and low-frequency subsequence, calculate the time interval between adjacent peaks to obtain the periodic sequence; calculate the average value of the periodic sequence, which is recorded as the average period.

[0021] As a preferred technical solution, different models are selected according to the fluctuation forms of different subsequences to predict carbon prices, specifically:

[0022] The autocorrelation function and partial autocorrelation function are used to select the best lag from the high-frequency subsequence or the low-frequency subsequence as input, and the augmented Dickey Fuller test is used to observe whether the high-frequency subsequence or the low-frequency subsequence is stable under the best lag value. If it is not stable, the difference or other stabilization processing is performed again;

[0023] The bidirectional long short-term memory model and BP network optimized by genetic algorithm are used to predict the high-frequency subsequence and low-frequency subsequence after decomposition respectively;

[0024] The prediction results of the high-frequency subsequence and the low-frequency subsequence are superimposed to obtain the carbon price range prediction results.

[0025] As a preferred technical solution, the different DLUBE neural network models are all obtained by optimizing parameters using genetic algorithms, specifically:

[0026] Determine the DLUBE neural network model, and use the parameter transfer strategy to initialize the DLUBE neural network model to obtain the initial population;

[0027] Decode and update the weights and biases of the DLUBE neural network model;

[0028] Use the pre-prepared training samples to train the DLUBE neural network model, and use the pre-prepared test samples to test the DLUBE neural network model to obtain the test error;

[0029] Calculate the fitness function and select the chromosome with the highest fitness function for replication;

[0030] Perform crossover and mutation operations to obtain a new population;

[0031] The iteration is performed until the end condition is reached, and the optimal individual is selected as the optimal weight and bias of the DLUBE neural network model.

[0032] As a preferred technical solution, the parameter transfer strategy is used to initialize the DLUBE neural network model, specifically:

[0033] Construct a neural network model with only one output layer for point prediction;

[0034] Use the gradient descent method to train the neural network model and obtain the parameters of the neural network model;

[0035] The parameters of the neural network model are transferred to the DLUBE neural network for initialization, and the initial weights of the DLUBE neural network are set.

[0036] As a preferred technical solution, the fitness function selects an improved interval loss function as the output of the objective function; the fitness function adopts a function derived based on maximum likelihood function estimation;

[0037] The selection operation adopts random traversal sampling;

[0038] The crossover operation adopts simulated binary crossover.

[0039] As a preferred technical solution, the method further includes:

[0040] The forecast results of the combined carbon price range are evaluated, and the evaluation indicators include interval coverage, interval width and indicators considering coverage and average width.

[0041] As a preferred technical solution, the calculation formula for the interval coverage is:

[0042]

[0043] Where n is the number of data, U i and L i Represent the upper and lower bounds of the prediction interval respectively; C i is the situation that the i-th data falls into the interval, 1 represents falling into, and 0 represents not falling into; y i is the actual observed value;

[0044] The calculation formula for the interval width is:

[0045]

[0046] Where R represents the range of the high-frequency subsequence or the low-frequency subsequence;

[0047] The calculation formula for the index considering coverage and average width is:

[0048]

[0049] Among them, PICP is the interval coverage and AIw is the average width.

[0050] Compared with the prior art, the present invention has the following advantages and beneficial effects: the present invention significantly improves the accuracy and efficiency of prediction by introducing a decomposition-integrated combined interval prediction framework and using a meta-heuristic algorithm (genetic algorithm) to optimize model parameters. Compared with the prior art, the technical means and innovations of the present invention are specifically embodied in the following aspects:

[0051] 1. Adopting decomposition and integration framework: The present invention decomposes complex prediction tasks into several sub-problems for modeling through the method of decomposition and integration. This method utilizes the multi-dimensional learning ability of multiple sub-models on data features to avoid the overfitting problem that may occur when a single deep learning model faces small samples or complex data. This decomposition strategy effectively improves the generalization ability of the model and makes the prediction results more robust.

[0052] 2. Combined prediction to improve data-driven decision-making capabilities: In the prediction framework, the present invention uses a combined prediction method to dynamically select and weight different sub-model prediction results according to the actual form and characteristics of the data, thereby generating the best interval prediction. This method fully exploits the characteristics of the data, effectively avoids the limitations of a single model for specific data forms, and greatly improves the applicability and accuracy of the prediction.

[0053] 3. Innovative combination of transfer learning and meta-heuristic algorithms to improve optimization efficiency: On the one hand, transfer learning uses the weights and biases of the pre-trained point prediction model as the initialization parameters of the DLUBE (Deep Learning Uncertainty Bound Estimation) model to provide a good initial solution for the interval prediction model. Compared with random initialization, this method makes full use of the feature representations learned in the point prediction model and accelerates the convergence of the model. On the other hand, based on transfer learning, the model parameters are optimized through genetic algorithms; genetic algorithms use population search to quickly find the global optimal solution, avoiding the problem that traditional optimization methods are prone to falling into local optimality. This innovative design enables the model to obtain excellent optimization results in a relatively short time.

[0054] In summary, the present invention overcomes the problems of inaccurate prediction and low optimization efficiency faced by the prior art in interval prediction by decomposing an integrated framework, a loss function based on maximum likelihood estimation, a combined prediction strategy, and a combination of transfer learning and genetic algorithm, significantly improves the prediction accuracy and adaptability of the model, and provides a new technical path and theoretical support for interval prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0056] Figure 1 Schematic diagram of the process of a combined carbon price prediction method based on decomposition integration and BiLSTM model in an embodiment of the present invention.

[0057] Figure 2 The figure is a flow chart of optimizing a neural network model using a genetic algorithm according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application.

[0059] Reference to "embodiments" in this application means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments.

[0060] like Figure 1 As shown, the combined carbon price prediction method based on decomposition integration and BiLSTM model in this embodiment includes the following steps:

[0061] S1. Collect carbon price time series data from the carbon emission rights center over a period of time.

[0062] S2. Perform data cleaning on carbon price time series data.

[0063] S3. The variational mode decomposition method is used to decompose the cleaned carbon price time series data into high-frequency subsequences and low-frequency subsequences. At the same time, the Pearson correlation analysis method is used to analyze the correlation coefficients and average periods of the high-frequency subsequences and low-frequency subsequences to verify the decomposition results of the variational mode decomposition method.

[0064] S4. Different DLUBE neural network models are selected to predict carbon prices according to the fluctuation forms of different subsequences to obtain the combined carbon price range prediction results; among them, different DLUBE neural network models include BP model and BiLSTM model; different DLUBE neural network models are obtained by parameter optimization using genetic algorithm.

[0065] Furthermore, the data cleaning of this application adopts the maximum value normalization method, and the formula is:

[0066]

[0067] Among them, p * is the cleaned carbon price time series data, p is the carbon price time series data, min(p) is the minimum value in the carbon price time series data, and max(p) is the maximum value in the carbon price time series data.

[0068] Furthermore, this application introduces a decomposition integration strategy to decompose the cleaned carbon price time series data into low-frequency subsequences and high-frequency subsequences to represent the normal market evolution trend of carbon prices and random uncontrollable system factors respectively; the decomposition steps using the variational mode decomposition method (VMD) are:

[0069] First, the cleaned carbon price time series data is subjected to Hilport transform to calculate the orthogonal components of the signal to obtain an analytical signal; this step enables the signal to be represented in complex form to facilitate the processing of spectral information.

[0070] Then, the objective function of the variational mode decomposition method (VMD) is constructed based on the analytical signal, and the Lagrange multiplier method is used to construct the constrained optimization objective function. The core of VMD is to decompose the signal x(t) into several modes (IMFs):

[0071]

[0072] where u k (t) is the kth mode, and K is the number of decomposed modes. Therefore, the goal of VMD is to minimize the following objective function:

[0073]

[0074] where w k is the center frequency of each mode, * represents the convolution operation, δ t is the time derivative operator.

[0075] Then the Lagrange multiplier method is used to construct the constrained optimization problem, and the signal reconstruction constraint Incorporating the optimization goal, the optimized objective function is:

[0076]

[0077] where λ is the Lagrange multiplier and α is the equilibrium parameter.

[0078] Then the optimized objective function is iteratively solved by the alternating direction multiplier method (ADMM) to obtain several modes u k (t) and its corresponding center frequency w k .

[0079] Finally, several modes are classified according to the center frequency or spectrum. k (t) can be divided into high-frequency subsequence and low-frequency subsequence, through its center frequency w k Or spectral distribution classification: high-frequency subsequences correspond to larger w k Value; low-frequency subsequences correspond to smaller w k value.

[0080] Furthermore, in order to better understand the relationship between each component and the original sequence, the Pearson correlation analysis method is used to perform correlation analysis, analyze the correlation coefficients and average periods of high-frequency subsequences and low-frequency subsequences, and verify and support the effectiveness of VMD in separating high-frequency and low-frequency components. Specifically:

[0081] The Pearson correlation coefficient is used to analyze the correlation between the high-frequency subsequence and the low-frequency subsequence and the cleaned carbon price time series data, and the correlation coefficient is calculated. If the correlation coefficient is lower than the set coefficient threshold, the cleaned carbon price time series data is decomposed again. In this embodiment, scipy.stats in python is used to calculate the correlation coefficient.

[0082] Calculate the average period: The average period refers to the average time interval of the subsequence fluctuations; first detect the peaks and use the peak detection algorithm (using scipy.signal.find_peaks in python) to find the peak position of each high-frequency subsequence and low-frequency subsequence; then calculate the time interval between adjacent peaks to obtain the periodic sequence; finally, calculate the average value of the periodic sequence, which is recorded as the average period.

[0083] Furthermore, different models are selected for carbon price prediction according to the fluctuation forms of different subsequences, specifically:

[0084] S4.1. Use the autocorrelation function (ACF) and partial autocorrelation function (PACF) to select the best lag from the high-frequency subsequence or low-frequency subsequence as input, and use the augmented Dickey Fuller (ADF) test to observe whether the high-frequency subsequence or low-frequency subsequence is stable under the best lag value. If it is not stable, re-difference or other stabilization processing is performed to ensure the prediction accuracy of the subsequent model. The autocorrelation coefficient (ACF) is a measure of autocorrelation, which shows the ratio of the covariance of the time series with its own lagged value to the variance, indicating the correlation of the time series with itself at different lags. The partial autocorrelation function (PACF) is another measure of autocorrelation, similar to the ACF, but excludes the influence of other lags, which helps to determine the order of autocorrelation patterns in the time series.

[0085] S4.2. Use the bidirectional long short-term memory model and BP network optimized by genetic algorithm to predict the decomposed high-frequency subsequence (trend component) and low-frequency subsequence (random component) respectively.

[0086] Furthermore, the bidirectional long short-term memory model (LSTM model) and BP network in the present application are both obtained by parameter optimization using genetic algorithms, and the optimization process is:

[0087] S4.2.1. Determine the DLUBE neural network model and use the parameter transfer strategy to initialize the DLUBE neural network model to obtain the initial population.

[0088] Since the lower and upper bounds are usually located near their actual values, the neural network model for interval prediction problems is highly similar to that for point prediction problems. Inspired by the concept of transfer learning, a parameter transfer strategy is adopted to speed up the training process. The steps are:

[0089] First, a neural network with only one output is constructed for point prediction. Then, the point prediction neural network model is trained using the gradient descent method. It only takes a few seconds to obtain the parameters of the neural network. Since the only difference between the point prediction model and the interval prediction model is the number of output neurons, these parameters can be set as the initial weights of the DLUBE neural network. By pre-training the point prediction neural network and migrating its parameters to the DLUBE neural network model, its convergence speed can be significantly improved.

[0090] Specifically, the initial parameters [ω(1), b(1)] of the DLUBE model are set to be exactly the same as the parameters [ω(1), b(1)] of the point prediction model. The initial parameters [ω(2U), b(2U), ω(2L), b(2L)] of the DLUBE model are set to [ω(2)+ε, b(2)+ε, ω(2)+ε, b(2)+ε], where different ε represent different random values, and the value range changes with model training.

[0091] Individual coding uses binary coding. Each individual is a binary string, which consists of the link weights between the input layer and the hidden layer, the hidden layer bias, the link weights between the hidden layer and the output layer, the bias, etc. Each weight and bias is encoded using the initial value obtained by parameter passing.

[0092] S4.2.2. Decode and update the weights and biases of the DLUBE neural network model.

[0093] S4.2.3. Use the pre-prepared training samples to train the DLUBE neural network model, and use the pre-prepared test samples to test the DLUBE neural network model to obtain the test error.

[0094] S4.2.4. Calculate the fitness function and select the chromosome with the highest fitness function for replication.

[0095] The invention is to make the prediction interval coverage rate as large as possible and the coverage width as small as possible when the neural network model predicts the interval, so the improved interval loss function is selected as the output of the objective function; the fitness function adopts the function derived based on the maximum likelihood estimation. The selection operation adopts random traversal sampling.

[0096] S4.2.5. Perform crossover and mutation operations to obtain a new population.

[0097] In this embodiment, the crossover operation adopts simulated binary crossover (SBX). The SBX operator is a probability-based crossover strategy that randomly selects the gene positions of the parental chromosomes and generates the corresponding gene positions of the new individuals according to a certain probability distribution to maintain the diversity of the population. The SBX operator can simulate a variety of biological crossover methods, such as single-point crossover, multi-point crossover, etc., and therefore has greater flexibility and adaptability. The mutation operation refers to generating the number of mutant genes with a certain probability, and randomly selecting the mutated genes. If the code of the selected gene is 1, it becomes 0; otherwise, it becomes 1.

[0098] S4.2.6. Iterate until the end condition is reached to obtain the optimal weights and biases of the DLUBE neural network model.

[0099] In genetic algorithms, chromosomes are codes used to represent solutions to problems. In this context: the chromosome in this application refers to the encoded form of the weights and biases (i.e., the parameter set of the network) of a neural network model. A chromosome can represent the current state of a complete neural network model, i.e., all its weights and biases. In neural network optimization, an individual is usually a specific instance of the weights and biases of a neural network model, that is, the solution corresponding to a chromosome. Therefore, an "individual" is actually used to represent a specific neural network parameter configuration. The initial population is generated by a certain strategy (such as random initialization or parameter transfer strategy), and it contains several individuals (i.e., initial solution candidates). The initial population is gradually optimized through iterations of the genetic algorithm (selection, crossover, mutation, etc.), and the final optimal individual is the optimal result.

[0100] S4.3. Superimpose the prediction results of the high-frequency subsequence and the low-frequency subsequence to obtain the carbon price range prediction result.

[0101] Furthermore, the method further comprises the steps of:

[0102] S5. Evaluate the combined carbon price range prediction results. The evaluation indicators include interval coverage, interval width, and indicators that consider coverage and average width.

[0103] Among them, interval coverage (PI coverage probability, PICP) is a basic evaluation indicator used to evaluate the total probability that the actual sample falls into the prediction interval. Its calculation expression is as follows:

[0104]

[0105] Where n is the number of data, U i and L i Represent the upper and lower bounds of the prediction interval respectively; C i is the situation that the i-th data falls into the interval, 1 represents falling into, and 0 represents not falling into; y i is the actual observed value.

[0106] The interval width (PI normalized averaged width, PINAW) is used to describe the width of the interval, and its calculation formula is:

[0107]

[0108] Wherein, R represents the range interval of the high-frequency subsequence or the low-frequency subsequence.

[0109] In view of the conflict between interval coverage and interval width, and considering that the interval width should be as small as possible under the condition that PICP satisfies the confidence interval, in order to best ensure the validity of the prediction interval, an indicator Ratio that considers coverage and average width is designed. Ratio measures the PICP value and the average width (AIw), which measures PICP and width. The best solutions can obtain higher Ratio values ​​because they achieve higher PICP values ​​through narrow intervals; high PICP values ​​achieved using larger intervals will lead to lower Ratio values. In contrast to CWC (coverage-length-based cirterion), even if the PICP obtained by the solution is slightly smaller than the target PINC (confidence interval), if AIw is small enough, the Ratio metric will still show it as a good solution, and the calculation formula is:

[0110]

[0111] Among them, PICP is the interval coverage and AIw is the average width.

[0112] It should be noted that, for the sake of convenience, the aforementioned method embodiments are all expressed as a series of action combinations, but those skilled in the art should know that the present invention is not limited to the described order of actions, because according to the present invention, certain steps can be performed in other orders or simultaneously.

[0113] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0114] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A combined carbon price prediction method based on decomposition integration and BiLSTM model, characterized in that: The steps include: Collect carbon price time series data from the carbon emission rights center over a period of time; Perform data cleaning on carbon price time series data; The variational mode decomposition method is used to decompose the cleaned carbon price time series data into high-frequency subsequences and low-frequency subsequences. At the same time, the Pearson correlation analysis method is used to analyze the correlation coefficients and average periods of the high-frequency subsequences and low-frequency subsequences to verify the decomposition results of the variational mode decomposition method. Different DLUBE neural network models are selected for carbon price prediction according to the fluctuation forms of different subsequences to obtain a combined carbon price range prediction result; the different DLUBE neural network models include a BP model and a BiLSTM model; the different DLUBE neural network models are all obtained by optimizing parameters using a genetic algorithm.

2. The combined carbon price prediction method according to claim 1, characterized in that: The data cleaning adopts the maximum value normalization method.

3. The combined carbon price prediction method according to claim 1, characterized in that: The variational mode decomposition method is used to decompose the cleaned carbon price time series data into high-frequency subsequences and low-frequency subsequences, specifically: The cleaned carbon price time series data is used to calculate the orthogonal components of the signal through Hilport transformation to obtain the analytical signal; The objective function of the variational mode decomposition method is constructed based on the analytical signal, and the constrained optimization objective function is constructed using the Lagrange multiplier method; The optimized objective function is iteratively solved by the alternating direction multiplier method to obtain several modes and their corresponding center frequencies; Several modes are classified according to the center frequency or spectrum to obtain high-frequency subsequences and low-frequency subsequences.

4. The combined carbon price prediction method according to claim 1, characterized in that: The analysis of the correlation coefficient and average period of the high-frequency subsequence and the low-frequency subsequence is specifically as follows: The Pearson correlation coefficient is used to analyze the correlation between the high-frequency subsequence and the low-frequency subsequence and the cleaned carbon price time series data, and the correlation coefficient is calculated. If the correlation coefficient is lower than the set coefficient threshold, the cleaned carbon price time series data is decomposed again; Use the peak detection algorithm to find the peak position of each high-frequency subsequence and low-frequency subsequence, calculate the time interval between adjacent peaks to obtain the periodic sequence; calculate the average value of the periodic sequence, which is recorded as the average period.

5. The combined carbon price prediction method according to claim 1, characterized in that: The method of selecting different models for carbon price prediction according to the fluctuation forms of different subsequences is as follows: The autocorrelation function and partial autocorrelation function are used to select the best lag from the high-frequency subsequence or the low-frequency subsequence as input, and the augmented Dickey Fuller test is used to observe whether the high-frequency subsequence or the low-frequency subsequence is stable under the best lag value. If it is not stable, the difference or other stabilization processing is performed again; The bidirectional long short-term memory model and BP network optimized by genetic algorithm are used to predict the high-frequency subsequence and low-frequency subsequence after decomposition respectively; The prediction results of the high-frequency subsequence and the low-frequency subsequence are superimposed to obtain the carbon price range prediction results.

6. The combined carbon price prediction method according to claim 1, characterized in that: The different DLUBE neural network models are all obtained by optimizing parameters using genetic algorithms, specifically: Determine the DLUBE neural network model, and use the parameter transfer strategy to initialize the DLUBE neural network model to obtain the initial population; Decode and update the weights and biases of the DLUBE neural network model; Use the pre-prepared training samples to train the DLUBE neural network model, and use the pre-prepared test samples to test the DLUBE neural network model to obtain the test error; Calculate the fitness function and select the chromosome with the highest fitness function for replication; Perform crossover and mutation operations to obtain a new population; The iteration is performed until the end condition is reached, and the optimal individual is selected as the optimal weight and bias of the DLUBE neural network model.

7. The combined carbon price prediction method according to claim 6, characterized in that: The parameter transfer strategy is used to initialize the DLUBE neural network model, specifically: Construct a neural network model with only one output layer for point prediction; Use the gradient descent method to train the neural network model and obtain the parameters of the neural network model; The parameters of the neural network model are transferred to the DLUBE neural network for initialization, and the initial weights of the DLUBE neural network are set.

8. The combined carbon price prediction method according to claim 6, characterized in that: The fitness function selects an improved interval loss function as the output of the objective function; the fitness function adopts a function derived based on maximum likelihood function estimation; The selection operation adopts random traversal sampling; The crossover operation adopts simulated binary crossover.

9. The combined carbon price prediction method according to claim 1, characterized in that: The method further comprises: The forecast results of the combined carbon price range are evaluated, and the evaluation indicators include interval coverage, interval width and indicators considering coverage and average width.

10. The combined carbon price prediction method according to claim 9, characterized in that: The calculation formula of the interval coverage is: Among them, n is the number of data, U i and L i Represent the upper and lower bounds of the prediction interval respectively; C i is the situation that the i-th data falls into the interval, 1 represents falling into, and 0 represents not falling into; y i is the actual observed value; The calculation formula for the interval width is: Where R represents the range of the high-frequency subsequence or the low-frequency subsequence; The calculation formula for the index considering coverage and average width is: Among them, PICP is the interval coverage and AIw is the average width.

Images