A carbon emission prediction method and device

Abstract

This application provides a carbon emission prediction method, comprising: acquiring carbon emission data; inputting the carbon emission data into a pre-trained improved LSTM model for data analysis and processing to obtain carbon emission prediction results; wherein, the improved LSTM model includes a MobileNetV2 pre-order network and an LSTM post-order network, the main body of the MobileNetV2 pre-order network consists of 17 inverse residual units, and each unit of the LSTM post-order network consists of a forget gate, an input gate, and an output gate. This application designs an improved LSTM model for carbon emission prediction, increasing the data dimension and improving prediction accuracy.

Description

Technical Field

[0001] This application relates to the field of application software management, and more particularly to a carbon emission prediction method. This application also relates to a carbon emission prediction device. Background Technology

[0002] Traditional Long Short-Term Memory (LSTM) networks are sensitive to time series data, but due to their dimensionality limitations, they cannot account for other factors that affect data changes, cannot add dimensionality, and cannot account for the significant impact of other factors on carbon emission changes.

[0003] Currently, most domestic carbon emission prediction technologies employ linear or nonlinear theories. Linear theory-based prediction models are theoretically simple and easy to understand. However, this method can only train the model using historical data for the current road segment, without considering other influences. Therefore, the model's performance deteriorates significantly as the prediction time interval decreases. Nonlinear theory-based models decompose data into signals of different resolutions through data processing. Prediction algorithms are then applied to each of these decomposed signals, and the prediction results are synthesized to obtain the final prediction. This model has strong anti-interference capabilities, but it is computationally intensive and inefficient. Furthermore, a corresponding model needs to be built for each prediction segment. Therefore, when predicting massive amounts of data, a large number of models need to be built, and the time spent training these models is substantial.

[0004] Meanwhile, considering the regional carbon emission factors released by the National Development and Reform Commission, which have a broad scope, they are only applicable to carbon emission data across a wide range of provinces and cities nationwide, and can only serve as trend predictions, resulting in significant errors in the prediction results. Furthermore, because historical data from individual samples is not incorporated, the uniqueness of sample data is not taken into account, thus the accuracy of existing carbon emission detection methods is not high. Summary of the Invention

[0005] The purpose of this invention is to overcome the deficiencies in the prior art and provide a carbon emission prediction method. This application also relates to a carbon emission prediction device.

[0006] This application provides a carbon emission prediction method, including:

[0007] Obtain carbon emission data;

[0008] The carbon emission data is input into a pre-trained improved LSTM model for data analysis and processing to obtain carbon emission prediction results. The improved LSTM model includes a MobileNetV2 pre-processing network and an LSTM post-processing network. The main body of the MobileNetV2 pre-processing network consists of 17 inverse residual units, and each unit of the LSTM post-processing network consists of a forget gate, an input gate, and an output gate.

[0009] Optionally, the carbon emission data includes a time dimension and a date type dimension.

[0010] Optionally, the encoding of the date dimension includes: weekdays, weekends, statutory holidays, and statutory holidays + weekends.

[0011] Optionally, the training method for the improved LSTM model includes:

[0012] Historical carbon emission data is collected, and the historical carbon emission data is classified, encoded, and divided into training set and test set;

[0013] Configure the network parameters of the improved LSTM model and input the training set data to obtain the training results;

[0014] Analyze the training results. If the training results converge, the training ends; otherwise, adjust the network parameters and retrain.

[0015] Optionally, the collection of historical carbon emission data includes: checking the validity of the data, reviewing and correcting erroneous data, and replacing empty data with average values.

[0016] This application also provides a carbon emission prediction device, comprising:

[0017] The acquisition module is used to acquire carbon emission data;

[0018] The analysis module is used to input the carbon emission data into a pre-trained improved LSTM model for data analysis and processing to obtain carbon emission prediction results. The improved LSTM model includes a MobileNetV2 pre-processing network and an LSTM post-processing network. The main body of the MobileNetV2 pre-processing network consists of 17 inverse residual units, and each unit of the LSTM post-processing network consists of a forget gate, an input gate, and an output gate.

[0019] Optionally, the carbon emission data includes a time dimension and a date type dimension.

[0020] Optionally, the encoding of the date dimension includes: weekdays, weekends, statutory holidays, and statutory holidays + weekends.

[0021] Optionally, a training module is also included, which performs the following steps:

[0022] Historical carbon emission data is collected, and the historical carbon emission data is classified, encoded, and divided into training set and test set;

[0023] Configure the network parameters of the improved LSTM model and input the training set data to obtain the training results;

[0024] Analyze the training results. If the training results converge, the training ends; otherwise, adjust the network parameters and retrain.

[0025] Optionally, the collection of historical carbon emission data includes: checking the validity of the data, reviewing and correcting erroneous data, and replacing empty data with average values.

[0026] The advantages and beneficial effects of this application are as follows:

[0027] This application provides a carbon emission prediction method, comprising: acquiring carbon emission data; inputting the carbon emission data into a pre-trained improved LSTM model for data analysis and processing to obtain carbon emission prediction results; wherein, the improved LSTM model includes a MobileNetV2 pre-order network and an LSTM post-order network, the main body of the MobileNetV2 pre-order network consists of 17 inverse residual units, and each unit of the LSTM post-order network consists of a forget gate, an input gate, and an output gate. This application designs an improved LSTM model for carbon emission prediction, increasing the data dimension and improving prediction accuracy. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the carbon emission prediction process in this application.

[0029] Figure 2 This is a schematic diagram of the improved LSTM model structure in this application.

[0030] Figure 3 This is a schematic diagram of the training of the improved LSTM model in this application.

[0031] Figure 4 This is a schematic diagram of the prediction results in this application.

[0032] Figure 5 This is a schematic diagram of the carbon emission prediction device in this application. Detailed Implementation

[0033] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention.

[0034] The following are examples of specific implementation processes provided to illustrate the technical solutions to be protected in this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can implement this application by different technical means under the guidance of the concept of this application. Therefore, this application is not limited to the specific embodiments below.

[0035] This application provides a carbon emission prediction method, comprising: acquiring carbon emission data; inputting the carbon emission data into a pre-trained improved LSTM model for data analysis and processing to obtain carbon emission prediction results; wherein, the improved LSTM model includes a MobileNetV2 pre-order network and an LSTM post-order network, the main body of the MobileNetV2 pre-order network consists of 17 inverse residual units, and each unit of the LSTM post-order network consists of a forget gate, an input gate, and an output gate. This application designs an improved LSTM model for carbon emission prediction, increasing the data dimension and improving prediction accuracy.

[0036] Figure 1 This is a schematic diagram of the carbon emission prediction process in this application.

[0037] Please refer to Figure 1 As shown, the carbon emission prediction method described in this application obtains carbon emission prediction data by inputting the acquired carbon emission data into a pre-trained carbon emission prediction model and performing data analysis through the carbon emission prediction model. The method includes:

[0038] S101 Acquire carbon emission data; S102 Input the carbon emission data into a pre-trained improved LSTM model for data analysis and processing to obtain carbon emission prediction results.

[0039] Specifically, after the carbon emission data is collected, its validity is checked during the data preprocessing process. This validity check includes processing erroneous and empty data.

[0040] To reduce the impact of erroneous and empty data on the final result, the average value of the 24-hour period of the day can be used to replace all invalid data.

[0041] In addition, a date type dimension is added to the existing time-based data. The dataset is then categorized by label to obtain data label types. To ensure data usability, the date format data needs to be properly encoded, as shown in Table 1 below:

[0042] Table 1 Encoding Rules

[0043]

[0044]

[0045] Finally, the collected and processed carbon emission data is input into the improved LSTM model for data analysis and processing to obtain carbon emission prediction results.

[0046] The Long Short-Term Memory (LSTM) network or LSTM model used in this application is the main prediction network. However, due to the fact that LSTM is sensitive to time series, the input dimension is limited, and it cannot fully take into account other important factors that affect the data results.

[0047] The concept of this application is to modify the feature extraction layer of the LSTM network into a pre-sequence feature extraction network that mainly uses convolutional operations.

[0048] Currently, mainstream feature extraction networks include VGGNet, ResNet, and MobileNetV2. However, with the rapid development of neural network technology, VGGNet networks suffer from vanishing gradients as the network deepens, and their accuracy no longer meets ideal requirements. While ResNet effectively solves the vanishing and exploding gradient problems caused by network depth and can extract more representative features with significantly improved accuracy, it requires a large number of parameters and has a long training time. MobileNetV2, proposed by the Google team in 2018, offers higher accuracy than ResNet, has a smaller model size, and solves the problem that depthwise separable convolutional kernels cannot learn features.

[0049] The main units of an LSTM network are the forget gate, input gate, and output gate. The forget gate determines how much of the cell state from the previous time step needs to be retained in the current time step; the input gate determines how much of the network's input data from the current time step needs to be saved to the cell state; and the output gate controls how much of the current cell state needs to be output to the current output value.

[0050] Based on the above analysis, this application ultimately adopts a method of using MobileNetV2 as the preceding feature extraction network and LSTM as the following prediction network for carbon emission prediction, which can also be called M2LSTM network (referred to as the improved LSTM model in this application).

[0051] Figure 2 This is a schematic diagram of the improved LSTM model structure in this application.

[0052] Please refer to Figure 2 As shown, the improved LSTM model includes a MobileNetV2 pre-processor network and an LSTM post-processor network. Among them, Figure 2 Part (a) is the input data with time and date type dimensions; Figure 2 Part (b) is the MobileNetV2 network structure, which mainly consists of 17 anti-residual units; Figure 3 Part (c) is a single unit of the LSTM network, consisting of a forget gate, an input gate, and an output gate.

[0053] Figure 3 This is a schematic diagram of the training of the improved LSTM model in this application.

[0054] Please refer to Figure 3 As shown, S201 collects historical carbon emission data, classifies and encodes the historical carbon emission data, and divides it into training set and test set.

[0055] Historical carbon emission data was collected. This patent uses carbon emission data from a thermal power plant in Liaoning Province for nearly 6 years, collected at 24:00 every day, as the experimental dataset, totaling 54,560 data points (not shown).

[0056] Historical carbon emission data is preprocessed. During preprocessing, the validity of the data is checked, including handling erroneous and empty data. To reduce the impact of erroneous and empty data on the final results, the average value of the current day's data is used to replace all invalid data.

[0057] Based on the existing time dimension of the data, a date type dimension is added. The dataset is then categorized by label to obtain data label types. To ensure data usability, the date format data needs to be properly encoded, and the encoding rules are shown in Table 1 above.

[0058] The data is divided appropriately. To ensure sufficient training and effective testing of the deep learning model, this patent uses 80% of the dataset as the training set and 20% as the test set.

[0059] Please refer to Figure 3 As shown, S202 sets the network parameters of the improved LSTM model and inputs the training set data to obtain the training results.

[0060] Configure the M2LSTM network parameters. The parameters that need to be set are shown in Table 2 below:

[0061] Table 2 Network Parameters

[0062]

[0063]

[0064] Input data with data labels into the improved LSTM model for training.

[0065] In this application, the maximum number of iterations is set to 300 training epochs, and the initial learning rate is set to 0.01. The learning rate decay method is to halve the learning rate and continue training if the accuracy does not improve for 3 consecutive epochs. If the model does not show significant optimization improvement for 10 consecutive epochs, early stopping is implemented. During training, the training set is split into two parts: 80% of the training data is used for model training, and the remaining 20% ​​of the data is used for model validation in each epoch.

[0066] The model file is saved once per training cycle. The input independent variables are carbon emission data and date-type data, totaling two sets. The output dependent variable is carbon emission data, totaling one set. The legacy duration is set to 30 for each time series, meaning that the next data is predicted using 30 data points.

[0067] Please refer to Figure 3 As shown, S203 analyzes the training results. If the training results converge, the training ends; otherwise, the network parameters are adjusted and the training is repeated.

[0068] The following is a display of the training results of the improved LSTM model described in this application.

[0069] Figure 4 This is a schematic diagram of the prediction results in this application.

[0070] Please refer to Figure 4 As shown, curve 301 represents the training and test values ​​of this application, and curve 302 represents the actual values ​​of the data in this application. The improved LSTM model built in this patent achieves a high level of carbon emission prediction, with the predicted curve closely matching the actual curve. Statistically, the average prediction accuracy for the integer part in the experimental dataset reaches 91.4%.

[0071] This application also provides a carbon emission prediction device, including: an acquisition module 401 and an analysis module 402.

[0072] Figure 5 This is a schematic diagram of the carbon emission prediction device in this application.

[0073] Please refer to Figure 5 As shown, module 401 acquires carbon emission data.

[0074] Specifically, after the carbon emission data is collected, its validity is checked during the data preprocessing process. This validity check includes processing erroneous and empty data.

[0075] To reduce the impact of erroneous and empty data on the final result, the average value of the 24-hour period of the day can be used to replace all invalid data.

[0076] In addition, a date type dimension is added to the existing time-based data. The dataset is then categorized by label to obtain data label types. To ensure data usability, the date format data needs to be properly encoded, as shown in Table 1 below:

[0077] Table 1 Encoding Rules

[0078]

[0079] The analysis module 402 inputs the carbon emission data into a pre-trained improved LSTM model for data analysis and processing to obtain carbon emission prediction results.

[0080] Finally, the collected and processed carbon emission data is input into the improved LSTM model for data analysis and processing to obtain carbon emission prediction results.

[0081] The Long Short-Term Memory (LSTM) network or LSTM model used in this application is the main prediction network. However, due to the fact that LSTM is sensitive to time series, the input dimension is limited, and it cannot fully take into account other important factors that affect the data results.

[0082] The concept of this application is to modify the feature extraction layer of the LSTM network into a pre-sequence feature extraction network that mainly uses convolutional operations.

[0083] This application uses MobileNetV2 as the preceding feature extraction network and LSTM as the following prediction network to predict carbon emissions, which can also be called M2LSTM network (referred to as the improved LSTM model in this application).

[0084] Figure 2 This is a schematic diagram of the improved LSTM model structure in this application.

[0085] Please refer to Figure 2 As shown, the improved LSTM model includes a MobileNetV2 pre-processor network and an LSTM post-processor network. Among them, Figure 2 Part (a) is the input data with time and date type dimensions; Figure 2 Part (b) is the MobileNetV2 network structure, which mainly consists of 17 anti-residual units; Figure 3 Part (c) is a single unit of the LSTM network, consisting of a forget gate, an input gate, and an output gate.

[0086] The carbon emission testing device further includes a training module, which performs the following steps:

[0087] Please refer to Figure 3 As shown, S201 collects historical carbon emission data, classifies and encodes the historical carbon emission data, and divides it into training set and test set.

[0088] Historical carbon emission data was collected. This patent uses carbon emission data from a thermal power plant in Liaoning Province for nearly 6 years, collected at 24:00 every day, as the experimental dataset, totaling 54,560 data points (not shown).

[0089] Historical carbon emission data is preprocessed. During preprocessing, the validity of the data is checked, including handling erroneous and empty data. To reduce the impact of erroneous and empty data on the final results, the average value of the current day's data is used to replace all invalid data.

[0090] Based on the existing time dimension of the data, a date type dimension is added. The dataset is then categorized by label to obtain data label types. To ensure data usability, the date format data needs to be properly encoded, and the encoding rules are shown in Table 1 above.

[0091] The data is divided appropriately. To ensure sufficient training and effective testing of the deep learning model, this patent uses 80% of the dataset as the training set and 20% as the test set.

[0092] Please refer to Figure 3 As shown, S202 sets the network parameters of the improved LSTM model and inputs the training set data to obtain the training results.

[0093] Configure the M2LSTM network parameters. The parameters that need to be set are shown in Table 2 below:

[0094] Table 2 Network Parameters

[0095]

[0096]

[0097] Input data with data labels into the improved LSTM model for training.

[0098] In this application, the maximum number of iterations is set to 300 training epochs, and the initial learning rate is set to 0.01. The learning rate decay method is to halve the learning rate and continue training if the accuracy does not improve for 3 consecutive epochs. If the model does not show significant optimization improvement for 10 consecutive epochs, early stopping is implemented. During training, the training set is split into two parts: 80% of the training data is used for model training, and the remaining 20% ​​of the data is used for model validation in each epoch.

[0099] The model file is saved once per training cycle. The input independent variables are carbon emission data and date-type data, totaling two sets. The output dependent variable is carbon emission data, totaling one set. The legacy duration is set to 30 for each time series, meaning that the next data is predicted using 30 data points.

[0100] Please refer to Figure 3 As shown, S203 analyzes the training results. If the training results converge, the training ends; otherwise, the network parameters are adjusted and the training is repeated.

Claims

1. A carbon emission prediction method, characterized in that, include: Obtain carbon emission data; The carbon emission data is input into a pre-trained improved LSTM model for data analysis and processing to obtain carbon emission prediction results. The improved LSTM model includes a MobileNetV2 pre-processing network and an LSTM post-processing network. The main body of the MobileNetV2 pre-processing network consists of 17 inverse residual units, and each unit of the LSTM post-processing network consists of a forget gate, an input gate, and an output gate.

2. The carbon emission prediction method according to claim 1, characterized in that, The carbon emission data includes a time dimension and a date type dimension.

3. The carbon emission prediction method according to claim 2, characterized in that, The encoding of the date type dimension includes: weekdays, weekends, statutory holidays, and statutory holidays + weekends.

4. The carbon emission prediction method according to claim 1, characterized in that, The training method for the improved LSTM model includes: Historical carbon emission data is collected, and the historical carbon emission data is classified, encoded, and divided into training set and test set; Configure the network parameters of the improved LSTM model and input the training set data to obtain the training results; Analyze the training results. If the training results converge, the training ends; otherwise, adjust the network parameters and retrain.

5. The carbon emission prediction method according to claim 1, characterized in that, The collection of historical carbon emission data includes: checking the validity of the data, reviewing and correcting erroneous data, and replacing empty data with average values.

6. A carbon emission prediction device, characterized in that, include: The acquisition module is used to acquire carbon emission data; The analysis module is used to input the carbon emission data into a pre-trained improved LSTM model for data analysis and processing to obtain carbon emission prediction results. The improved LSTM model includes a MobileNetV2 pre-processing network and an LSTM post-processing network. The main body of the MobileNetV2 pre-processing network consists of 17 inverse residual units, and each unit of the LSTM post-processing network consists of a forget gate, an input gate, and an output gate.

7. The carbon emission prediction device according to claim 6, characterized in that, The carbon emission data includes a time dimension and a date type dimension.

8. The carbon emission prediction device according to claim 7, characterized in that, The encoding of the date type dimension includes: weekdays, weekends, statutory holidays, and statutory holidays + weekends.

9. The carbon emission prediction device according to claim 6, characterized in that, It also includes a training module, which performs the following steps: Historical carbon emission data is collected, and the historical carbon emission data is classified, encoded, and divided into training set and test set; Configure the network parameters of the improved LSTM model and input the training set data to obtain the training results; Analyze the training results. If the training results converge, the training ends; otherwise, adjust the network parameters and retrain.

10. The carbon emission prediction device according to claim 6, characterized in that, The collection of historical carbon emission data includes: checking the validity of the data, reviewing and correcting erroneous data, and replacing empty data with average values.

Images