Carbon price prediction method based on empirical mode decomposition and multi-time-sequence prediction model integration
By adopting empirical modal decomposition and multi-time sequence prediction model integration methods in carbon price prediction, combined with causal convolution network and other models, the problem of failure to fully consider power market factors and data preprocessing improvement space in the existing technology is solved, and a more efficient carbon price prediction effect is achieved.
Patent Information
- Application Number
- CN202411758979.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-12-19
- Filing Date
- 2024-12-03
- Publication Date
- 2025-03-04
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing carbon price prediction methods fail to fully consider power market factors, and there is room for improvement in data preprocessing and model training, resulting in poor carbon price prediction results.
Using the method based on empirical modal decomposition (EMD) and multi-time sequence prediction model integration, through data decomposition and dimensionality reduction, the causal convolution network, GRU, LSTM and Transformer models are trained, and the results are integrated for prediction.
By taking into account power market factors and multi-model integration, the accuracy of carbon price prediction is improved, the uncertainty that may arise from individual models is reduced, and more abundant carbon price range prediction information is provided.
Smart Images
Figure CN119558902A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of carbon price prediction, and in particular to a carbon price prediction method based on empirical mode decomposition and multi-time series prediction model integration. Background Art
[0002] In the carbon market, companies can trade carbon quotas, and the carbon price refers to the price of carbon quotas in the carbon trading process. Carbon price forecasting is conducive to the development of the carbon market and brings more profits to companies. In carbon price forecasting, artificial intelligence technology can effectively extract data features and make accurate forecasts.
[0003] Among the existing methods, there are predictions using the quantile LASSO model and XGBoost. These methods treat the data at each time point as independent data and train at each time point that is considered to be independently distributed. Since the time series characteristics of the carbon price series are not used, the prediction effect is difficult to guarantee. Other methods take time series characteristics into consideration and use models that can extract time series, such as RNN and LSTM. These methods treat the carbon price series as a time series data and input it into the time series model for training.
[0004] It can be seen that most of the existing methods do not take into account the impact of electricity market factors, and there is room for improvement in data preprocessing and model training. Summary of the invention
[0005] The purpose of the present invention is to provide a carbon price forecasting method based on empirical mode decomposition and multi-time series forecasting model integration. By using multiple models and integrating the results for forecasting, the uncertainty that may be generated by a single model can be avoided. The combination of multiple models can eliminate certain uncertainties and solve the difficult problem of unstable and highly random carbon price series in carbon price forecasting. Moreover, considering the impact of the electricity market on carbon, the model training is more comprehensive and complete.
[0006] To achieve the above object, the present invention provides a carbon price prediction method based on empirical mode decomposition and multi-time series prediction model integration, comprising the following steps:
[0007] S1. Data collection: Collect historical carbon price data and carbon price impact data sets in the target area, and organize the carbon price data and carbon price impact data sets into carbon price series and carbon price impact series groups respectively;
[0008] S2. Data preprocessing: Use EMD decomposition method to decompose the carbon price series and perform PCA dimensionality reduction. At the same time, perform PCA dimensionality reduction on the carbon price impact data group, and combine the carbon price series after PCA dimensionality reduction and the carbon price impact series group;
[0009] S3, training model: randomly input the combined data of step S2 and the carbon price sequence into the causal convolutional network model, GRU model, LSTM model and Transformer model for training respectively;
[0010] S4. Input the carbon price impact data group preprocessed by step S2 into the causal convolutional network model, GRU model, LSTM model and Transformer model trained in step S3 respectively, and combine the output results of the causal convolutional network model, GRU model, LSTM model and Transformer model as the final prediction result.
[0011] Preferably, in step S1, the carbon valence sequence is recorded as Y = {y1, y2, ... y n}, where y1,y2,…y n Respectively represent the carbon prices at the 1st, 2nd, 3rd, …, nth time nodes;
[0012] The carbon price impact data group described in step S1 includes an energy data group and a power data group, and the carbon price impact sequence group obtained by sorting out the p groups of carbon price impact data collected is recorded as {X1, X2, ..., X i ,…,X p},in Indicates n rows m i Matrix of columns, m i Indicates that the total number of carbon price impact data collected for the i-th group is m i dimension.
[0013] Preferably, step S2 specifically includes the following steps:
[0014] S21. Use EMD decomposition method to analyze the carbon valence sequence Y = {y1, y2, ... y n} is decomposed and the output is recorded as {IMF1,IMF2,…,IMF i ,…,IMF m}, where IMF i ∈R n*1 , R n*1 Represents a matrix with n rows and 1 column;
[0015] S22, use PCA to analyze the output results {IMF1, IMF2, ..., IMF i ,…,IMF m} to reduce the dimension and output the matrix [F1, F2, …, F m' ], and the contribution rate of each dimension is recorded as e1, e2,…, e m' ;
[0016] S23, use PCA to analyze the carbon price impact sequence groups {X1, X2, …, X i ,…,X p} to reduce the dimension and output p dimensionality reduction results {X'1,X'2,…X' p},in Indicates n rows m i ' Column matrix, m i ' represents the dimension of the carbon price impact data of the i-th group after dimensionality reduction;
[0017] S24, the output of step S22 [F1, F2, ..., F m' ] and {X'1, X'2, ... X' output in step S23 p} to obtain m' combination data:
[0018] [F1,X'1,X'2,…X′ p ], [F2,X'1,X'2,…X′ p ], …, [F m' ,X'1,X'2,…X′ p ].
[0019] Preferably, the causal convolutional network model described in step S3 includes an input layer, a temporal network layer, a fully connected layer for carbon price prediction, a fully connected layer for upper limit prediction of carbon price prediction interval, and a fully connected layer for lower limit prediction of carbon price prediction interval, which are arranged in sequence;
[0020] The loss function expression of the fully connected layer for carbon price prediction is as follows:
[0021] Loss=(y pre -y truth )
[0022] Where Loss represents the loss of the fully connected layer used for carbon price prediction; y pre represents the predicted carbon price; y truth represents the true label, y truth =Y;
[0023] The loss function expressions of the fully connected layer for predicting the lower limit of the carbon price prediction interval and the fully connected layer for predicting the upper limit of the carbon price prediction interval are as follows:
[0024] loss1=max{-α×(y pre1 -y truth ),(1-α)×(y pre1 -y truth )}
[0025] loss2=max{(α-1)×(y pre2 -y truth ),α×(ypre2 -y truth )}
[0026] Where loss1 and loss2 represent the losses of the fully connected layer for predicting the lower limit of the carbon price prediction interval and the fully connected layer for predicting the upper limit of the carbon price prediction interval, respectively; α represents the quantile level; y pre1 and pre2 They represent the lower and upper limits of the carbon price forecast interval respectively;
[0027] The m' combination data obtained in step S2 are used as input vectors, and the carbon price sequence Y is used as the true label. A total of m' causal convolutional network models, GRU models, LSTM models, and Transformer models are trained, denoted as M1, M2, ..., M m' .
[0028] Preferably, in step S4, the carbon price impact data of the m' group k days ago are input into M1, M2, ..., M m' , and the prediction results are After PCA dimensionality reduction, the contribution rates e1, e2, …, e m' , perform weighted average on the results to get the final prediction result.
[0029] Therefore, the carbon price prediction method based on empirical mode decomposition and multi-time series prediction model integration of the present invention has the following beneficial effects:
[0030] (1) The present invention specifically considers the synergy between electricity and carbon prices, introduces electricity market factors into carbon price forecasting, uses EMD and PCA to process data, groups data to train multiple forecasting models, and improves the accuracy of forecasting results through integrated voting;
[0031] (2) In the data processing of the present invention, the carbon price series is decomposed by using EMD and the PCA is used for dimensionality reduction, and then the model is trained on each component separately and the results are integrated; the present invention takes into account more carbon-related factors, and the method can easily expand more carbon-related data;
[0032] (3) The present invention uses multiple models to predict the integrated results, which can avoid the uncertainty that may be generated by a single model; the combination of multiple models can eliminate certain uncertainties;
[0033] (4) The present invention can predict the carbon price range, provide users with richer information, and help users make decisions.
[0034] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 is a flow chart of an embodiment of the present invention;
[0036] Figure 2 Graphs showing interval prediction results of the causal convolutional network model of an embodiment of the present invention; (a) is an example graph showing the upper limit of the carbon price prediction interval; (b) is an example graph showing the lower limit of the carbon price prediction interval; (c) is an example graph showing carbon price prediction;
[0037] Figure 3 This is a European carbon price series data diagram according to an embodiment of the present invention;
[0038] Figure 4 This is a result diagram after EMD decomposition according to an embodiment of the present invention. DETAILED DESCRIPTION
[0039] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.
[0040] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.
[0041] A carbon price prediction method based on empirical mode decomposition and multi-time series prediction model integration includes the following steps:
[0042] S1. Data collection: Collect historical carbon price data and carbon price impact data sets in the target area, and organize the carbon price data and carbon price impact data sets into carbon price series and carbon price impact series groups respectively;
[0043] In step S1, the carbon valence sequence is recorded as Y = {y1, y2, ... y n}, where y1,y2,…y n Respectively represent the carbon prices at the 1st, 2nd, 3rd, …, nth time nodes;
[0044] The carbon price impact data group described in step S1 includes an energy data group and a power data group, and the carbon price impact sequence group obtained by sorting out the p groups of carbon price impact data collected is recorded as {X1, X2, ..., X i ,…,X p},in Indicates n rows m i Matrix of columns, m i Indicates that the total number of carbon price impact data collected for the i-th group is m i dimension.
[0045] In this embodiment, the energy data group is used to reflect the carbon emission intensity of different energy types and the progress of clean energy transformation: such as the increase in wind and solar energy production capacity, changes in fossil fuel prices and their potential impact on the energy consumption structure; the power data group is used to monitor and analyze the decarbonization process of the power industry, the changing trend of power demand, and the development of power storage technology.
[0046] S2. Data preprocessing: Use EMD decomposition method to decompose the carbon price series and perform PCA dimensionality reduction. At the same time, perform PCA dimensionality reduction on the carbon price impact data group, and combine the carbon price series after PCA dimensionality reduction and the carbon price impact series group;
[0047] Step S2 specifically includes the following steps:
[0048] S21. Use EMD decomposition method to analyze the carbon valence sequence Y = {y1, y2, ... y n} is decomposed and the output is recorded as {IMF1,IMF2,…,IMF i ,…,IMF m}, where IMF i ∈R n*1 , R n*1 Represents a matrix with n rows and 1 column;
[0049] S22, use PCA to analyze the output results {IMF1, IMF2, ..., IMF i ,…,IMF m} to reduce the dimension and output the matrix [F1, F2, …, F m' ], and the contribution rate of each dimension is recorded as e1, e2,…, e m' ;
[0050] S23, use PCA to analyze the carbon price impact sequence groups {X1, X2, …, X i ,…,X p} to reduce the dimension and output p dimensionality reduction results {X'1,X'2,…X' p},in Indicates n rows m i ' Column matrix, m i ' represents the dimension of the carbon price impact data of the i-th group after dimensionality reduction;
[0051] S24, the output of step S22 [F1, F2, ..., F m' ] and {X'1, X'2, ... X' output in step S23 p} to obtain m' combination data:
[0052] [F1,X'1,X'2,…X′ p], [F2,X'1,X'2,…X′ p ], …, [F m' ,X'1,X'2,…X p ].
[0053] S3, training model: randomly input the combined data of step S2 and the carbon price sequence into the causal convolutional network model, GRU model, LSTM model and Transformer model for training respectively;
[0054] The causal convolutional network model described in step S3 includes an input layer, a temporal network layer, a fully connected layer for carbon price prediction, a fully connected layer for upper limit prediction of carbon price prediction interval, and a fully connected layer for lower limit prediction of carbon price prediction interval, which are arranged in sequence;
[0055] The loss function expression of the fully connected layer for carbon price prediction is as follows:
[0056] Loss=(y pre -y truth )
[0057] Where Loss represents the loss of the fully connected layer used for carbon price prediction; y pre represents the predicted carbon price; y truth represents the true label, y truth =Y;
[0058] The loss function expressions of the fully connected layer for predicting the lower limit of the carbon price prediction interval and the fully connected layer for predicting the upper limit of the carbon price prediction interval are as follows:
[0059] loss1=max{-α×(y pre1 -y truth ),(1-α)×(y pre1 -y truth )}
[0060] loss2=max{(α-1)×(y pre2 -y truth ),α×(y pre2 -y truth )}
[0061] Where loss1 and loss2 represent the losses of the fully connected layer for predicting the lower limit of the carbon price prediction interval and the fully connected layer for predicting the upper limit of the carbon price prediction interval, respectively; α represents the quantile level; y pre1 and pre2 They represent the lower and upper limits of the carbon price forecast interval respectively;
[0062] like Figure 2As shown in the figure, the output of the prediction interval is realized through the above causal convolutional network model. For example, the causal convolutional network model predicts that the carbon price is 5.34, the prediction interval is [4.23, 6.73], and the confidence level is 90%. This means that the causal convolutional network model is 90% sure that the carbon price on the next day will fall in the interval [4.23, 6.73]. The interval prediction can provide more information for market participants and assist decision makers in formulating trading strategies.
[0063] The m' combination data obtained in step S2 are used as input vectors, and the carbon price sequence Y is used as the true label. A total of m' causal convolutional network models GRU model, LSTM model, and Transformer model are trained, denoted as M1, M2, ..., M m' , that is, a total of m' causal convolutional network models, GRU models, LSTM models and Transformer models are obtained.
[0064] S4. Input the carbon price impact data group preprocessed by step S2 into the causal convolutional network model, GRU model, LSTM model and Transformer model trained in step S3 respectively, and combine the output results of the causal convolutional network model, GRU model, LSTM model and Transformer model as the final prediction result.
[0065] In step S4, the carbon price impact data of group m' k days ago are input into M1, M2, ..., M m' , and the prediction results are After PCA dimensionality reduction, the contribution rates e1, e2, …, e m' , perform weighted average on the results to get the final prediction result.
[0066] Let’s take the European carbon price series as an example. Figure 3 As shown, a total of 3351 days of data were collected; model prediction was performed using the prediction method of the present invention, as shown in FIG. Figure 4 As shown in the figure, it is the result after EMD decomposition, with a total of 9 components; the final prediction result is as follows Figure 2 As shown in (c), it can be seen that the prediction result is very good and the broken line fitting is normal, thus confirming the effectiveness of the present invention.
[0067] Therefore, the present invention provides a carbon price forecasting method based on empirical mode decomposition and multi-time series forecasting model integration, which uses multiple models and integrates the results for forecasting, thereby avoiding the uncertainty that may be caused by a single model; the combination of multiple models can eliminate certain uncertainties and solve the difficult problems of unstable and highly random carbon price series in carbon price forecasting; and considering the impact of the electricity market on carbon, the model training is more comprehensive and complete; at the same time, by predicting the carbon price range, it provides users with richer information to help them make decisions.
[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.
Claims
1. A carbon price prediction method based on empirical mode decomposition and multi-time series prediction model integration, characterized by: The following steps are involved: S1. Data collection: Collect historical carbon price data and carbon price impact data sets in the target area, and organize the carbon price data and carbon price impact data sets into carbon price series and carbon price impact series groups respectively; S2. Data preprocessing: Use EMD decomposition method to decompose the carbon price series and perform PCA dimensionality reduction. At the same time, perform PCA dimensionality reduction on the carbon price impact data group, and combine the carbon price series after PCA dimensionality reduction and the carbon price impact series group; S3, training model: randomly input the combined data of step S2 and the carbon price sequence into the causal convolutional network model, GRU model, LSTM model and Transformer model for training respectively; S4. Input the carbon price impact data group preprocessed by step S2 into the causal convolutional network model, GRU model, LSTM model and Transformer model trained in step S3 respectively, and combine the output results of the causal convolutional network model, GRU model, LSTM model and Transformer model as the final prediction result.
2. The carbon price prediction method based on empirical mode decomposition and multi-time series prediction model integration according to claim 1 is characterized by: In step S1, the carbon valence sequence is recorded as Y = {y1, y2, ... y n }, where y1,y2,…y n Respectively represent the carbon prices at the 1st, 2nd, 3rd, …, nth time nodes; The carbon price impact data group described in step S1 includes an energy data group and a power data group, and the carbon price impact sequence group obtained by sorting out the p groups of carbon price impact data collected is recorded as {X1, X2, ..., X i ,…,X p },in Indicates n rows m i Matrix of columns, m i Indicates that the total number of carbon price impact data collected for the i-th group is m i dimension.
3. The carbon price prediction method based on empirical mode decomposition and multi-time series prediction model integration according to claim 2 is characterized in that: Step S2 specifically includes the following steps: S21. Use EMD decomposition method to analyze the carbon valence sequence Y = {y1, y2, ... y n } is decomposed and the output is recorded as {IMF1,IMF2,…,IMF i ,…,IMF m }, where IMF i ∈R n*1 , R n*1 Represents a matrix with n rows and 1 column; S22, use PCA to analyze the output results {IMF1, IMF2, ..., IMF i ,…,IMF m } to reduce the dimension and output the matrix [F1, F2, …, F m' ], and the contribution rate of each dimension is recorded as e1, e2,…, e m' ; S23, use PCA to analyze the carbon price impact sequence groups {X1, X2, …, X i ,…,X p } to reduce the dimension and output p dimensionality reduction results {X'1,X'2,…X' p },in Indicates n rows m i ' Column matrix, m i ' represents the dimension of the carbon price impact data of the i-th group after dimensionality reduction; S24, the output of step S22 [F1, F2, ..., F m' ] and {X'1, X'2, ... X' output in step S23 p } to obtain m' combination data: [F1,X′1,X′2,…X′ p ]、[F2,X′1,X′2,…X′ p ]、…、[F m' ,X′1,X′2,…X′ p ]。 4. The carbon price prediction method based on empirical mode decomposition and multi-time series prediction model integration according to claim 3 is characterized by: The causal convolutional network model described in step S3 includes an input layer, a temporal network layer, a fully connected layer for carbon price prediction, a fully connected layer for upper limit prediction of carbon price prediction interval, and a fully connected layer for lower limit prediction of carbon price prediction interval, which are arranged in sequence; The loss function expression of the fully connected layer for carbon price prediction is as follows: Loss=(y pre -y truth ) Where Loss represents the loss of the fully connected layer used for carbon price prediction; y pre represents the predicted carbon price; y truth represents the true label, y truth =Y; The loss function expressions of the fully connected layer for predicting the lower limit of the carbon price prediction interval and the fully connected layer for predicting the upper limit of the carbon price prediction interval are as follows: loss1=max{-α×(y pre1 -and truth ),(1-α)×(and pre1 -and truth )} loss2=max{(α-1)×(y pre2 -and truth ),α×(and pre2 -and truth )} Where loss1 and loss2 represent the losses of the fully connected layer for predicting the lower limit of the carbon price prediction interval and the fully connected layer for predicting the upper limit of the carbon price prediction interval, respectively; α represents the quantile level; y pre1 and pre2 They represent the lower and upper limits of the carbon price forecast interval respectively; The m' combination data obtained in step S2 are used as input vectors, and the carbon price sequence Y is used as the true label. A total of m' causal convolutional network models, GRU models, LSTM models, and Transformer models are trained, denoted as M1, M2, ..., M m' .
5. The carbon price prediction method based on empirical mode decomposition and multi-time series prediction model integration according to claim 4 is characterized in that: In step S4, the carbon price impact data of group m' k days ago are input into M1, M2, ..., M m' , and the prediction results are After PCA dimensionality reduction, the contribution rates e1, e2, …, e m' , perform weighted average on the results to get the final prediction result.