Carbon emission prediction method based on weighted correction composite quantile regression minimum gating unit

By combining convolutional neural networks and composite quantile regression minimum gating units, the shortcomings of the existing technology in complex nonlinear relationships and large-scale data processing are solved, and more accurate and robust carbon emission prediction is achieved, especially In terms of long-term prediction and probability prediction.

CN119005401APending Publication Date: 2024-11-22CHANGCHUN UNIV OF TECH
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411023986.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

Existing carbon emission prediction models have shortcomings in dealing with complex nonlinear relationships and large-scale data, especially in long-term prediction and feature extraction. Deep learning models such as LSTM improve prediction accuracy but have gradient disappearance or explosion problems, and The quality of probability prediction depends on the feature extraction ability of historical data.

Method used

Combining convolutional neural network (CNN) and composite quantile regression minimum gating unit (CQRMGU), through weighted correction of composite quantile regression, quantile prediction is converted into density prediction using kernel density estimation, providing a more comprehensive carbon Emission change probability prediction.

Benefits of technology

Improve the accuracy and robustness of carbon emission prediction, extract local key information through the CNN layer, learn time correlation of the MGU layer, output conditional quantiles for point prediction, interval prediction and probability density prediction, avoiding model overfitting and cross phenomenon enhances the overall performance of prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119005401A_ABST
    Figure CN119005401A_ABST
Patent Text Reader

Abstract

The invention discloses a carbon emission prediction method based on a weighted correction composite quantile regression minimum gating unit, which combines quantile regression with a neural network through combination of a convolutional neural network and the composite quantile regression minimum gating unit. For a quantile crossing phenomenon, weighted correction improvement is carried out on a quantile curve, and a penalty term is added on a loss function to avoid model overfitting. According to the model, local key information is effectively extracted through the convolutional neural network layer, correlation of the local key information in time is fully learned by using the minimum gating unit layer, conditional quantiles are output for prediction, a final prediction result is obtained, and the prediction result is evaluated through an evaluation index. The prediction combined with the existing quantile regression and neural network has a better effect, and the accuracy of carbon emission prediction is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of carbon emission prediction, and in particular to a prediction method based on a weighted modified composite quantile regression minimum gating unit. Background Art

[0002] In the current context of increasing global attention to climate change and sustainable development, climate change caused by carbon emissions has become a major challenge facing the world today. Long-term predictions of carbon dioxide emissions provide a basis for the formulation and implementation of emission reduction policies, while short-term predictions are very important for grasping the dynamic trends of carbon dioxide emissions. At present, research on carbon dioxide emission prediction involves multiple fields, and is mainly divided into annual emission predictions and daily emission predictions according to data types. The former is regarded as a long-term prediction, while the latter is regarded as a short-term prediction. Carbon emission prediction research is mainly divided into three types of models: statistical models, machine learning models, and deep learning models. Each type has different characteristics in terms of model construction and performance. Statistical models are good at simulating historical trends and periodicity. However, they often have shortcomings when faced with complex nonlinear relationships and the influence of multiple factors, especially in long-term predictions and processing large-scale data. In contrast, machine learning models show higher prediction accuracy and robustness. With the continuous increase in data volume and problem complexity, deep learning models have gradually gained attention. Models such as long short-term memory (LSTM) use their powerful nonlinear modeling capabilities and sensitivity to time dependence to make them suitable for large-scale, high-dimensional time series data. Deep learning models have been proven to be good at extracting potential patterns and improving prediction accuracy, but there is a problem of gradient disappearance or explosion. The quality of probabilistic forecasting also depends on the ability to extract features from historical data, which can be done with the help of point forecasting models. Unlike deterministic forecasting, probabilistic forecasting has attracted more and more research attention due to its strong ability to capture the uncertainty of future demand. Probabilistic forecasting can better describe the uncertainty components contained in deterministic forecasting results through three forms: prediction interval, quantile, and probability density function. Summary of the invention

[0003] In view of the above problems, the present invention provides a carbon emission prediction method based on the minimum gating unit of weighted modified composite quantile regression, which combines the deep learning model with the composite quantile regression, and provides more prediction information by using the deep quantile regression carbon emission forecast under different quantile conditions. Kernel density estimation (KDE) is a bridge connecting density prediction and quantile prediction. By converting quantile prediction into density prediction, the probability density result can provide a comprehensive perspective on carbon emission changes.

[0004] In order to achieve the above object, the present invention adopts the following technical solution.

[0005] A carbon emission prediction method based on weighted modified composite quantile regression minimum gating unit includes the following steps:

[0006] Step S1, obtain carbon emission target data, perform data preprocessing on the data set, including normalization and date feature expansion and one-hot encoding;

[0007] Step S2, dividing the preprocessed data set and setting the number of input data batches;

[0008] Step S3, using a convolutional neural network (CNN) to perform preliminary feature extraction on the data, and then using the composite quantile regression (CQR) minimum gated unit (MGU) to predict the conditional quantile;

[0009] Step S4: Apply the trained CQRMGU model to the test set to obtain the prediction results, and evaluate the prediction results using evaluation indicators.

[0010] Specifically, in step S1, the data is normalized using the Max-Min method to convert it into the interval [0,1], and the normalization formula is:

[0011]

[0012] The date feature expansion is to expand the year, month, and day data to days of the week, working days, and holidays, and perform one-hot encoding as feature input.

[0013] Specifically, in step S2, the preprocessed data set is divided into a training set and a test set in a ratio of 8:2; and the batch size is set to 128.

[0014] Specifically, in step S3, a convolutional neural network (CNN) is used to perform preliminary feature extraction on the data. CNN is usually composed of a convolution layer, a pooling layer, and a fully connected layer. X is the input of the convolution layer. A one-dimensional convolution kernel is used to slide from left to right on X to set different weights for data points to extract key information. The specific method for predicting conditional quantiles using CQRMGU is:

[0015] (1) Input training set data

[0016] (2) Calculate the forget gate f of MGU t , candidate memory cells and the hidden layer output value h t

[0017] f t =σ(W f [h t-1 ,xt ]+b f ) (2)

[0018]

[0019]

[0020] Among them, W f , W h 、b f 、b h They represent the corresponding weight matrix and bias vector respectively, σ(·) is the Sigmoid function, and tanh(·) is the hyperbolic tangent function.

[0021] The final output of the output layer:

[0022]

[0023] Where W s is the connection weight matrix between the hidden layer and the output layer, b s represents the corresponding bias vector.

[0024] (3) Based on the composite quantile regression loss function and to avoid overfitting of the model, a penalty term is added to the objective function to obtain the training model.

[0025]

[0026]

[0027] In the formula, K represents the number of quantiles.

[0028] (4) Correct the crossover phenomenon. Assume that the quantile τ1 < τ2, and the quantile regression curve and They intersect at point Q, where the horizontal coordinate is x Q If the curve to be corrected is Select a τ3 value that is close to τ2, and τ2<τ3, then α and β satisfy the equation group:

[0029]

[0030] Then the corrected τ2 regression equation is:

[0031]

[0032] In step S4, the trained CQRMGU model is applied to the test set to obtain a prediction result, which is characterized in that the prediction result is evaluated by an evaluation index. The prediction result is evaluated by three evaluation indicators: point prediction interval prediction and probability density prediction, specifically:

[0033] Point prediction evaluation indicators:

[0034] Mean absolute error MAE, root mean square error RMSE, mean absolute percentage error MAPE

[0035]

[0036]

[0037]

[0038] Evaluation indicators for interval prediction:

[0039] There are three evaluation indicators: prediction interval coverage probability (PICP), prediction interval average width (PIAW) and Winkler score (WS). When the confidence level is α, PICP measures the percentage of prediction intervals that contain observed values.

[0040]

[0041]

[0042] PIAW is a sharpness index used to evaluate the average width of PI.

[0043]

[0044] In order to evaluate the applicability of interval prediction, it is necessary to ensure the minimum interval width under a certain interval coverage.

[0045]

[0046]

[0047] Probabilistic prediction evaluation indicators:

[0048] The continuous ranking probability score (CRPS) has a good ability to evaluate the comprehensive performance of probability prediction. The smaller the CRPS, the better the probability prediction result. The formula of CRPS is:

[0049]

[0050]

[0051]

[0052] Among them, p(y i ) represents y i The probability density function of is the probability model obtained from the predicted values, is a unit step function.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] The carbon emission probability prediction method of the present invention combines the convolutional neural network (CNN) and the composite quantile regression (CQR) minimum gated unit (MGU) to combine quantile regression and neural network to address the problems existing in the existing neural network model. In view of the quantile crossing phenomenon, the quantile curve is weighted and corrected, and a penalty term is added to the loss function to avoid overfitting of the model. The model effectively extracts local key information through the CNN layer, uses the MGU layer to fully learn the temporal correlation of local key information, and outputs conditional quantiles for point prediction, interval prediction and probability density prediction. The prediction combined with the existing quantile regression and neural network has a better effect and improves the accuracy of carbon emission prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is the overall flow chart of the present invention;

[0056] Figure 2 This is the MGU structure diagram; DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] Example

[0059] like Figure 1 As shown, the present invention provides a carbon emission prediction method based on weighted modified composite quantile regression minimum gating unit, comprising the following steps:

[0060] Step S1, obtain carbon emission target data, perform data preprocessing on the data set, including normalization and date feature expansion and one-hot encoding;

[0061] Step S2, dividing the preprocessed data set and setting the number of input data batches;

[0062] Step S3: Use convolutional neural network (CNN) to perform preliminary feature extraction on the data, and then use composite quantile regression (CQR) minimum gated unit (MGU) to predict conditional quantiles, and perform weighted correction to avoid crossover;

[0063] Step S4: Apply the trained CQRMGU model to the test set to obtain the probability prediction results, and evaluate the prediction results through evaluation indicators.

[0064] Specifically, in step S1, the data is normalized using the Max-Min method to convert it into the interval [0,1], and the normalization formula is:

[0065]

[0066] The date feature expansion is to expand the year, month, and day data to weeks, weekdays, and holidays, and perform one-hot encoding as feature input. The encoding results are shown in Table 1 below.

[0067] Table 1 One-hot encoding

[0068] Specifically, in step S2, the preprocessed data set is divided into a training set and a test set in a ratio of 8:2; and the batch size is set to 128.

[0069] Specifically, in step S3, a convolutional neural network (CNN) is used to perform preliminary feature extraction on the data. CNN is usually composed of a convolution layer, a pooling layer, and a fully connected layer. X is the input of the convolution layer. A one-dimensional convolution kernel is used to slide from left to right on X to set different weights for data points to extract key information. The specific method for predicting conditional quantiles using CQRMGU is:

[0070] (1) Input training set data

[0071] (2) Figure 2 As shown, calculate the forget gate f of MGU t , candidate memory cells and the hidden layer output value h t

[0072] f t =σ(W f [h t-1 ,x t ]+b f ) (2)

[0073]

[0074]

[0075] Among them, W f , W h 、b f 、b h They represent the corresponding weight matrix and bias vector respectively, σ(·) is the Sigmoid function, and tanh(·) is the hyperbolic tangent function.

[0076] The final output of the output layer:

[0077]

[0078] Where W s is the connection weight matrix between the hidden layer and the output layer, b s represents the corresponding bias vector.

[0079] (3) Based on the composite quantile regression loss function and to avoid overfitting of the model, a penalty term is added to the objective function to obtain the training model.

[0080]

[0081]

[0082] In the formula, K represents the number of quantiles.

[0083] (4) Correct the crossover phenomenon. Assume that the quantile τ1 < τ2, and the quantile regression curve and They intersect at point Q, where the horizontal coordinate is x Q If the curve to be corrected is Select a τ3 value that is close to τ2, and τ2<τ3, then α and β satisfy the equation group:

[0084]

[0085] Then the corrected τ2 regression equation is:

[0086]

[0087] In step S4, the trained CQRMGU model is applied to the test set to obtain a prediction result, which is characterized in that the prediction result is evaluated by an evaluation index. The prediction result is evaluated by three evaluation indicators: point prediction interval prediction and probability density prediction, specifically:

[0088] Point prediction evaluation indicators:

[0089] Mean absolute error MAE, root mean square error RMSE, mean absolute percentage error MAPE

[0090]

[0091]

[0092]

[0093] Evaluation indicators for interval prediction:

[0094] There are three evaluation indicators: prediction interval coverage probability (PICP), prediction interval average width (PIAW) and Winkler score (WS). When the confidence level is α, PICP measures the percentage of prediction intervals that contain observed values.

[0095]

[0096]

[0097] PIAW is a sharpness index used to evaluate the average width of PI.

[0098]

[0099] In order to evaluate the applicability of interval prediction, it is necessary to ensure the minimum interval width under a certain interval coverage.

[0100]

[0101]

[0102] Probabilistic prediction evaluation indicators:

[0103] The continuous ranking probability score (CRPS) has a good ability to evaluate the comprehensive performance of probability prediction. The smaller the CRPS, the better the probability prediction result. The formula of CRPS is:

[0104]

[0105]

[0106]

[0107] Among them, p(y i ) represents y i The probability density function of is the probability model obtained from the predicted values, is a unit step function.

[0108] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A carbon emission prediction method based on weighted modified composite quantile regression minimum gating unit, It is characterized in that The following steps are involved: Step S1, obtain carbon emission target data, perform data preprocessing on the data set, including normalization and date feature expansion and one-hot encoding; Step S2, dividing the preprocessed data set and setting the number of input data batches; Step S3: Use convolutional neural network (CNN) to perform preliminary feature extraction on the data, and then use composite quantile regression (CQR) minimum gated unit (MGU) to predict conditional quantiles, and perform weighted correction to avoid crossover; Step S4: Apply the trained CQRMGU model to the test set to obtain the prediction results, and evaluate the prediction results using evaluation indicators.

2. According to claim 1, a carbon emission prediction method based on weighted modified composite quantile regression minimum gating unit, It is characterized in that In step S1, the data is normalized using the Max-Min method to convert it into the interval [0,1]. The normalization formula is: The date feature expansion is to expand the year, month, and day data to days of the week, working days, and holidays, and perform one-hot encoding as feature input.

3. According to claim 1, a carbon emission prediction method based on weighted modified composite quantile regression minimum gating unit, It is characterized in that In step S2, the preprocessed data set is divided into a training set and a test set in a ratio of 8:2; the batch size is set to 128.

4. According to claim 1, a carbon emission prediction method based on weighted modified composite quantile regression minimum gating unit, It is characterized in that In step S3, a convolutional neural network (CNN) is used to perform preliminary feature extraction on the data. CNN is usually composed of a convolution layer, a pooling layer, and a fully connected layer. X is the input of the convolution layer. A one-dimensional convolution kernel is used to slide from left to right on X to set different weights for data points to extract key information. The specific method of using CQRMGU to predict conditional quantiles is: (1) Input training set data (2) Calculate the forget gate f of MGU t , candidate memory cells and the hidden layer output value h t f t =σ(W f [h t-1 ,x t ]+b f ) (2) Among them, W f , W h , b f , b h They represent the corresponding weight matrix and bias vector respectively, σ(·) is the Sigmoid function, and tanh(·) is the hyperbolic tangent function. The final output of the output layer: Where W s is the connection weight matrix between the hidden layer and the output layer, b s represents the corresponding bias vector. (3) Based on the composite quantile regression loss function and to avoid overfitting of the model, a penalty term is added to the objective function to obtain the training model. In the formula, K represents the number of quantiles. (4) Correct the crossover phenomenon. Let the quantile τ 1 <τ 2 , quantile regression curve and They intersect at point Q, where the horizontal coordinate is x Q If the curve to be corrected is Select and τ 2 The closer τ 3 value, and τ 2 <τ 3 , then α and β satisfy the equation group: Then the corrected τ 2 The regression equation is:

5. According to claim 1, a carbon emission prediction method based on weighted modified composite quantile regression minimum gating unit, It is characterized in that In step S4, the trained CQRMGU model is applied to the test set to obtain a prediction result, which is characterized in that the prediction result is evaluated by an evaluation index. The prediction result is evaluated by three evaluation indicators: point prediction interval prediction and probability density prediction, specifically: Point prediction evaluation indicators: Mean absolute error MAE, root mean square error RMSE, mean absolute percentage error MAPE Evaluation indicators for interval prediction: There are three evaluation indicators: prediction interval coverage probability (PICP), prediction interval average width (PIAW) and Winkler score (WS). When the confidence level is α, PICP measures the percentage of prediction intervals that contain observed values. PIAW is a sharpness index used to evaluate the average width of PI. In order to evaluate the applicability of interval prediction, it is necessary to ensure the minimum interval width under a certain interval coverage. Probabilistic prediction evaluation indicators: The continuous ranking probability score (CRPS) has a good ability to evaluate the comprehensive performance of probability prediction. The smaller the CRPS, the better the probability prediction result. The formula of CRPS is: Among them, p(y i ) represents y i The probability density function of is the probability model obtained from the predicted values, is a unit step function.

Citation Information

Cited By

  • Low-carbon construction evaluation method based on data processing

    CN120373665A