An annual carbon emission prediction method, a prediction device, a medium and an electronic device

Abstract

The present application provides a kind of annual carbon emission prediction method, prediction device, medium and electronic equipment, comprising the following steps: according to the data set obtained by preset condition processing, at least including carbon emission information and the data subset of different years is obtained;Analysis of the data subset, form the visual view of the different years and the carbon emission information, and form the relationship evaluation matrix for estimating the carbon emission information and the different years;Carbon emission prediction model including deep learning model and integrated model is constructed, when the data subset is input into the carbon emission prediction model, carbon emission trend prediction result close to real data is obtained.The present application combines deep learning model and integrated model to predict annual carbon emission, compared with existing prediction model, the prediction result is more accurate, and the prediction result is closer to real carbon emission data.

Description

Technical Field

[0001] This invention relates to the field of data prediction technology, and in particular to an annual carbon emission prediction method, prediction device, medium, and electronic equipment. Background Technology

[0002] With the increasing severity of global climate change, reducing greenhouse gas emissions has become a global focus. As a major component of greenhouse gases, accurate prediction of carbon emissions is crucial for developing effective emission reduction policies and measures. In recent years, the rapid development of big data analytics and artificial intelligence technologies has provided new methods and tools for carbon emission prediction. Carbon emission prediction is fundamental to evaluating and optimizing emission reduction strategies. By accurately predicting future carbon emission trends, businesses and research institutions can better set emission reduction targets, optimize resource allocation, and assess policy effectiveness. For example, businesses can optimize production processes and reduce their carbon footprint by predicting carbon emissions.

[0003] Currently, carbon emission prediction methods include correlation analysis algorithms and model estimation methods. Correlation analysis algorithms predict the carbon emissions of a given emitter by analyzing the relationship between carbon emissions and strongly correlated factors (such as energy consumption and industrial production indices); model estimation methods predict carbon emissions by constructing mathematical models.

[0004] For example, a carbon emission prediction model for the Yellow River Basin, which uses a Long Short-Term Memory (LSTM) network optimized by the Sparrow Search Algorithm (SSA), can accurately predict the carbon emissions of a specific region (Yellow River Basin). However, this model only focuses on regional predictions, and SSA-LSTM is not applicable to corporate emissions. The imputation method based on ensemble machine learning enhances the prediction method of agricultural carbon dioxide emissions. It has high accuracy in predicting CO2 emissions in agricultural datasets. However, the drawback is that this prediction method only focuses on data imbalance and imputation, rather than time dependence. A machine learning ensemble model for predicting the CO2 footprint of GGBFS-based geopolymer concrete can achieve high-precision prediction of CO2 in concrete production. However, it has limitations and is only able to predict the emissions of specific materials. The hybrid short-term carbon emission prediction model for China's aviation industry highlights the potential of hybrid models to significantly improve the accuracy of carbon emission prediction in the aviation sector. It has superior prediction performance in short-term aviation emissions. Its shortcomings are that it can only predict for specific industries and focuses on short-term prediction. The method of coordinating extreme learning machines and information algorithms for carbon emission prediction has proven effective in handling large datasets and improving the model's generalization ability in carbon emission prediction, and can reduce the error rate of carbon emission prediction. However, it can only be optimized for specific carbon datasets and lacks deep learning components for temporal analysis.

[0005] Therefore, how to provide a carbon emission prediction method that can at least make the carbon emission prediction results more accurate and closer to the actual carbon emission data is a technical problem that urgently needs to be solved. Summary of the Invention

[0006] To address the shortcomings mentioned above, this invention provides an annual carbon emission forecasting method, device, medium, and electronic equipment that provides more accurate forecasts that are closer to actual carbon emission data.

[0007] To achieve the above objectives, in a first aspect, the present invention provides an annual carbon emission prediction method, comprising the following steps: The dataset is processed according to preset conditions to obtain a subset of data that includes at least carbon emission information and data from different years. Analyze the data subset to form a visual diagram of the carbon emission information for different years, and form an evaluation matrix for estimating the relationship between the carbon emission information and the different years; A carbon emission prediction model is constructed that includes a deep learning model and an ensemble model. When the data subset is input into the carbon emission prediction model, a carbon emission trend prediction result that closely approximates the real data is obtained.

[0008] In one embodiment, the process of processing the dataset according to preset conditions to obtain a subset of data that includes at least carbon emission information and different years includes the following steps: The obtained dataset is filtered according to the first preset condition, and data that is related to the first preset condition is deleted to form the first dataset; The first dataset is sorted according to the second preset condition to form a sequence corresponding to the second preset condition; The sequence is filtered according to a third preset condition to form a subset of data that does not contain the third preset condition.

[0009] In one embodiment, the process of establishing a carbon emission prediction model and inputting the carbon emission information into the model to obtain a carbon emission trend prediction result that approximates the actual data includes the following steps: Construct a carbon emission prediction model that includes both deep learning models and ensemble models; The data subsets are respectively input into the deep learning model to obtain multiple first predicted carbon emission data; At least one of the first predicted carbon emission data is input into the integrated model to obtain a second predicted carbon emission data that is close to the actual carbon emission data. The second predicted carbon emission data is evaluated to determine its accuracy.

[0010] In one embodiment, the deep learning model includes a multilayer perceptron, a recurrent neural network, and a long short-term memory neural network, wherein: After the data subset is input into the multilayer sensor, the first predicted carbon emission data is obtained using the following formula: yˆ=f(WL···f(Wif(Wix+bi)+bi)···+bL), Where yˆ is the first predicted carbon emission data, f is the activation function, Wi and bi are the weights and biases of the i-th layer, i is a number, and L is the layer number; After the data subset is input into the recurrent neural network, the first predicted carbon emission data is obtained sequentially using the following formula. ht=f(Wxxt+Whht 1+b), Among them, h t Let be the hidden state vector at time t, f be the activation function, Wx, Wh, and b be the learning functions, xt be the input vector at time t, and h be the learning vector. t -1 represents the hidden state at the previous time step, and b is the bias vector; yˆt=g(Wyht+c), Where yˆt represents the first predicted carbon emission data at time t, g is the activation function, Wy and c are the learning functions, and h t Let be the hidden state vector at time t; After the data subset is input into the long short-term memory neural network, the first predicted carbon emission data is obtained sequentially using the following formula. ft=σ(Wf[ht 1, xt]+bf), Where ft is the forget gate vector, σ is the Sigmoid activation function, Wf is the weight matrix of the forget gate, and [ht] [1, xt] is the concatenation vector, and bf is the bias vector of the forget gate; it=σ(Wi[ht 1, xt]+bi), Where it is the input gate vector, σ is the Sigmoid activation function, Wi is the input gate weight matrix, and [ht] [1, xt] is the concatenation vector, and bi is the input gate bias vector; ct=ft⊙ct 1+it⊙tanh(Wc[ht 1, xt]+bc), Where ct is the cell state at the current time step, ft is the forgetting gate vector, and ct 1 represents the cell state at the previous time step, it represents the input gate vector, and tanh(Wc[ht]) represents the input gate vector. 1, xt]+bc) represents the candidate cell state, where: Where Wc is the weight matrix, [ht] 1, xt] represents the hidden state ht from the previous time step. 1. Concatenate the vectors with the current input xt, bc is the bias term, tanh is the hyperbolic tangent activation function, and ⊙ is the Hadamard element-wise multiplication operator; ot=σ(Wo[ht 1, xt]+bo), Where ot is the output gate vector, σ is the Sigmoid activation function, Wo is the output gate weight matrix, and [ht] [1, xt] represents vector concatenation, and bo is the output gate bias vector; ht = ot⊙tanh(ct), Where ht is the hidden state at the current time step, ot is the output gate vector, ⊙ is the Hadamard element-wise multiplication operator, tanh is the element-wise function, and ct is the vector.

[0011] In one embodiment, the integrated model is divided into a first integrated model, a second integrated model and a third integrated model, wherein the first integrated model is formed by integrating the multilayer perceptron, the recurrent neural network and the long short-term memory neural network, and is used to obtain the average value of the first predicted carbon emission data; The second integrated model is formed by integrating two of the multilayer perceptron, the recurrent neural network, and the long short-term memory neural network, and is used to obtain the average value of the first predicted carbon emission data; The third integrated model is formed by integrating multiple multilayer perceptrons, or multiple recurrent neural networks, or multiple long short-term memory neural networks, and is used to obtain the average value of the first predicted carbon emission data.

[0012] In one embodiment, the third integration model is a single-mode integration model or an interleaved integration model, wherein: The single-mode ensemble model is formed by integrating multiple multilayer perceptrons with different training parameters, or multiple recurrent neural networks with different training parameters, or multiple long short-term memory neural networks with different training parameters. The interleaved ensemble model is formed by integrating multiple multilayer perceptrons with different training times, or multiple recurrent neural networks with different training times, or multiple long short-term memory neural networks with different training times.

[0013] Secondly, the present invention also provides a prediction apparatus for implementing the above-described annual carbon emission prediction method, comprising: The data processing module is used to process the obtained dataset according to preset conditions to obtain a data subset that includes at least carbon emission information and data from different years. The data analysis module is used to analyze the data subset, form a visual representation of the carbon emission information for different years, and form an evaluation matrix for estimating the relationship between the carbon emission information and the different years. A carbon emission prediction model is used to construct a carbon emission prediction model that includes a deep learning model and an ensemble model. When the data subset is input into the carbon emission prediction model, a carbon emission trend prediction result that closely approximates the real data is obtained.

[0014] In one embodiment, the carbon emission prediction module includes a construction module, a deep learning model, an ensemble model, and an evaluation model, wherein: The building module is used to construct a carbon emission prediction model that includes deep learning and ensemble models; The deep learning model is used to derive multiple first predicted carbon emission data based on the data subset; The integrated model receives at least one of the first predicted carbon emission data as input to derive a second predicted carbon emission data that closely approximates the actual carbon emission data. The evaluation model is used to evaluate the second predicted carbon emission data and determine the accuracy of the second predicted carbon emission data.

[0015] Thirdly, the present invention also provides a medium storing a computer program that, when executed by a processor, implements the steps of the above-described annual carbon emission prediction method.

[0016] Fourthly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described annual carbon emission prediction method.

[0017] Compared with the prior art, the present invention has one of the following advantages: This invention combines a deep learning model with an ensemble model for annual carbon emission prediction. Compared to existing prediction models, the prediction results are more accurate and closer to real carbon emission data. In addition, the combination of deep learning and ensemble models can capture nonlinear relationships and time patterns, integrate nonlinear interactions and time dynamics, and handle cross-industry enterprise-level emissions and short-term and long-term predictions. It has enhanced temporal modeling and improved cross-departmental generalization capabilities, providing a general solution for enterprise carbon emissions. When the second ensemble model is chosen to be formed by integrating a multilayer perceptron with a recurrent neural network, a balance between accuracy and model stability can be achieved, suggesting that simpler and well-tuned models may sometimes outperform complex architectures in CO2 emission prediction tasks. Attached Figure Description

[0018] Figure 1 This is a flowchart of the annual carbon emission prediction method in this invention; Figure 2 A correlation chart of carbon emissions over the years; Figure 3 A graph showing the carbon emissions of a selected facility over time; Figure 4 A chart showing the year-on-year percentage change in carbon emissions for selected facilities; Figure 5 A chart showing annual carbon emission forecasts using a deep learning model; Figures 6a to 6e A chart for predicting annual carbon emissions using an integrated model; Figure 7 A chart comparing the root mean square errors of all models; Figure 8 This is a structural block diagram of the prediction device in this invention; Figure 9 This is a structural block diagram of the electronic device in this invention.

[0019] The main reference numerals are as follows: 1-Data processing module; 2-Data analysis module; 3-Carbon emission prediction model; 4-Memory; 5-Communication interface; 6-Processor; 7-Bus system. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0021] In the description of this invention, it should be understood that the terms "upper," "lower," "top surface," "bottom surface," and "inner," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. It should be noted that when a component is considered to be "connected" to another component, it can be directly connected to the other component or there may be a component positioned centrally. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0022] Example 1 like Figure 1 As shown, this embodiment provides a method for predicting annual carbon emissions, including the following steps: S100. Process the obtained dataset according to preset conditions to obtain a data subset that includes at least carbon emission information and data from different years. S200: Analyze subsets of data to form a visual representation of carbon emission information for different years, and to form an assessment matrix for estimating the relationship between carbon emission information and different years; S300. Construct a carbon emission prediction model that includes a deep learning model and an ensemble model. When a subset of data is input into the carbon emission prediction model, a carbon emission trend prediction result that closely approximates the real data is obtained.

[0023] Specifically, step S100 includes the following steps: S101. The dataset obtained by filtering according to the first preset condition is used to delete data related to the first preset condition to form the first dataset; S102. Sort the first dataset according to the second preset condition to form a sequence corresponding to the second preset condition; S103. Filter the sequence according to the third preset condition to form a data subset that does not contain the third preset condition.

[0024] Furthermore, in step S100, the obtained dataset is the relevant carbon emission report.

[0025] Preferably, the carbon emission report may include information such as company information and emission information. The company information includes company details, company location, and postal code of the company location. The emission information includes year, facility type (e.g., facility registration ID, facility name, facility installation location), and emission type (e.g., carbon dioxide emission information, nitrous oxide emission information, hydrofluorocarbon emission information, stationary combustion data).

[0026] In step S101, content unrelated to emission information in the obtained dataset is filtered and deleted (i.e., the first preset condition), forming a first dataset that only includes emission information.

[0027] Preferably, company information in the obtained dataset is filtered and deleted to form a first dataset that includes only emissions information.

[0028] In step S102, emissions are grouped according to year and specified emission type (i.e., second preset condition) to facilitate clear annual trend analysis of emission types.

[0029] Preferably, the data is grouped according to year and carbon dioxide emission information to facilitate clear annual trend analysis of carbon dioxide.

[0030] In step S103, records with missing values ​​or zero emissions are excluded (i.e., the third preset condition) to avoid outliers and ensure the robustness of model training.

[0031] Through the above steps S101 to S103, the filtered dataset includes only important emission information, excluding information on indirect emitters and enterprises not within the scope, which facilitates centralized analysis.

[0032] Specifically, step S200 includes the following steps: S201. Use the processed carbon dioxide emission information to perform trend analysis and generate a visual chart of carbon dioxide emission information for different years. S202. Form an assessment matrix to estimate the relationship between carbon dioxide emission information and different years.

[0033] In step S201, for example, the steps of performing trend analysis and generating a visual chart are as follows: 1) Calculate the year-on-year annual increase in the data subset containing carbon dioxide emission information, and simultaneously calculate the carbon dioxide emission intensity and enterprise average emissions to provide indicators for subsequent multi-perspective comparisons. 2) Apply time series decomposition methods such as moving average or HP filter to break down carbon dioxide emissions into three parts: long-term trend, periodic fluctuation and random residual, identify the main trend and fluctuation range, and determine whether there is a plateau or inflection point. Choose a line chart to display the raw emissions and long-term trends for different years.

[0034] In step S202, the steps for forming the evaluation matrix are exemplarily as follows: 1) Define the matrix structure Lines: represent different years; Columns: Represent indicators related to carbon dioxide emissions (e.g., total carbon dioxide emissions, carbon dioxide intensity, etc.). Fill the matrix: Fill the corresponding cells with the annual carbon dioxide emission data to form the initial matrix; 2) Time series feature extraction Matrix profiling techniques are used to extract features from time series data to capture trends and periodic changes in carbon dioxide emission data. Dimensionality reduction: Principal component analysis and other methods are used to reduce data dimensionality and identify the main influencing factors. 3) Model building and prediction Use time series models to model and predict carbon dioxide emission data; The carbon dioxide data was divided into training and testing sets. The model was trained and the prediction accuracy was evaluated. The parameters were adjusted to optimize the model. 4) Anomaly detection, correction, and output By comparing the model's predicted values ​​with the actual values, the error is calculated and outlier data points are identified. Correct or remove the detected abnormal data; After ensuring the accuracy of the data, an evaluation matrix is ​​formed.

[0035] Specifically, step S300 includes the following sub-steps: S301. Construct a carbon emission prediction model by combining deep learning models with ensemble models; S302. Input a subset of data into a deep learning model to obtain multiple first-predicted carbon emission data; S303. Input at least one first predicted carbon emission data into the integrated model to obtain a second predicted carbon emission data that is close to the actual carbon emission data. S304. Evaluate the second predicted carbon emission data to determine the accuracy of the second predicted carbon emission data.

[0036] Furthermore, in step S301, a deep learning model is used to capture complex nonlinear temporal dependencies, and an ensemble model is used to fuse multi-source features with deep learning output to improve overall prediction accuracy.

[0037] Further, in step S302, the deep learning model is a multilayer perceptron (i.e., MLP), a recurrent neural network (i.e., RNN), and a long short-term memory neural network (i.e., LSTM), wherein: After a subset of data is input into the multilayer sensor, the first predicted carbon emission data is obtained using the following formula. yˆ=f(WL···f(Wif(Wix+bi)+bi)···+bL), Where yˆ is the first predicted carbon emission data, f is the activation function, Wi and bi are the weights and biases of the i-th layer, i is a number, and L is the layer number; After a subset of data is input into the recurrent neural network, the first predicted carbon emission data is obtained sequentially using the following formula. ht=f(Wxxt+Whht 1+b), Among them, h t Let be the hidden state vector at time t, f be the activation function, Wx, Wh, and b be the learning functions, xt be the input vector at time t, and h be the learning vector. t -1 represents the hidden state at the previous time step, and b is the bias vector; yˆt=g(Wyht+c), Where yˆt represents the first predicted carbon emission data at time t, g is the activation function, Wy and c are the learning functions, and h t Let be the hidden state vector at time t.

[0038] After a subset of data is input into a long short-term memory neural network, the first predicted carbon emission data is obtained sequentially using the following formulas. ft=σ(Wf[ht 1, xt]+bf), Where ft is the forget gate vector, σ is the Sigmoid activation function, Wf is the weight matrix of the forget gate, and [ht] [1, xt] is the concatenation vector, and bf is the bias vector of the forget gate; it=σ(Wi[ht 1, xt]+bi), Where it is the input gate vector, σ is the Sigmoid activation function, Wi is the input gate weight matrix, and [ht] [1, xt] is the concatenation vector, and bi is the input gate bias vector; ct=ft⊙ct 1+it⊙tanh(Wc[ht 1, xt]+bc), Where ct is the cell state at the current time step, ft is the forgetting gate vector, and ct 1 represents the cell state at the previous time step, it represents the input gate vector, and tanh(Wc[ht]) represents the input gate vector. 1, xt]+bc) represents the candidate cell state, where Wc is the weight matrix, [ht] is the weight matrix, and [ht] is the weight matrix. 1, xt] represents the hidden state ht from the previous time step. 1. Concatenate the vectors with the current input xt, bc is the bias term, tanh is the hyperbolic tangent activation function, and ⊙ is the Hadamard element-wise multiplication operator; ot=σ(Wo[ht 1, xt]+bo), Where ot is the output gate vector, σ is the Sigmoid activation function, Wo is the output gate weight matrix, and [ht] [1, xt] represents vector concatenation, and bo is the output gate bias vector; ht = ot⊙tanh(ct), Where ht is the hidden state at the current time step, ot is the output gate vector, ⊙ is the Hadamard element-wise multiplication operator, tanh is the element-wise function, and ct is the vector.

[0039] In step S302, the multilayer perceptron, recurrent neural network, and long short-term memory neural network are all trained with six years of carbon dioxide emission data to facilitate the prediction of carbon dioxide emissions in the following year.

[0040] Furthermore, in step S303, in order to improve the accuracy of the prediction, the ensemble model is divided into a first ensemble model, a second ensemble model and a third ensemble model. The first ensemble model (i.e., AE) is formed by integrating a multilayer perceptron, a recurrent neural network and a long short-term memory neural network, and is used to obtain the average value of the first predicted carbon emission data. The second ensemble model is formed by integrating two of the multilayer perceptron, recurrent neural network and long short-term memory neural network, and is used to derive the average value of the first predicted carbon emission data. The third ensemble model is formed by integrating multiple multilayer perceptrons, or multiple recurrent neural networks, or multiple long short-term memory neural networks, and is used to derive the average value of the first predicted carbon emission data.

[0041] Furthermore, after inputting the first predicted carbon emission data obtained from the multilayer perceptron, recurrent neural network, and long short-term memory neural network into the first ensemble model (i.e., AE), the average value of the first predicted carbon emission data can be obtained using the following formula: y^`=1 / 3(y^MLP+y^RNN+y^LSTM), Wherein, y^` is the average value of the first predicted carbon emission data, y^MLP is the first predicted carbon emission data obtained through a multilayer perceptron, y^RNN is the first predicted carbon emission data obtained through a recurrent neural network, and y^LSTM is the first predicted carbon emission data obtained through a long short-term memory neural network.

[0042] Furthermore, preferably, among the multilayer perceptron, recurrent neural network, and long short-term memory neural network, the multilayer perceptron and recurrent neural network are the two best-performing models. Therefore, the second ensemble model is formed by integrating the multilayer perceptron and the recurrent neural network. After the first predicted carbon emission data obtained by the multilayer perceptron and the recurrent neural network are input into the second ensemble model (i.e., MR), the average value of the first predicted carbon emission data is obtained using the following formula: y^`=1 / 2(y^MLP+y^RNN), Where y^` is the average value of the first predicted carbon emission data, y^MLP is the first predicted carbon emission data obtained by the multilayer perceptron, and y^RNN is the first predicted carbon emission data obtained by the recurrent neural network.

[0043] Furthermore, the third ensemble model can be either a single-mode ensemble model or an interleaved ensemble model, which can increase the diversity of the third ensemble model. Specifically: a single-mode ensemble model is a single-mode multilayer perceptron model formed by ensembled multiple multilayer perceptrons with different training parameters (i.e., MLPe), a single-mode recurrent neural network model formed by ensembled multiple recurrent neural networks with different training parameters (i.e., RNNe), or a single-mode long short-term memory neural network model formed by ensembled multiple long short-term memory neural networks with different training parameters (i.e., LTMe). An interleaved ensemble model is an interleaved multilayer perceptron model formed by ensembled multiple multilayer perceptrons with different training times (i.e., MLPes), an interleaved recurrent neural network model formed by ensembled multiple recurrent neural networks with different training times (i.e., RNNes), or an interleaved long short-term memory neural network model formed by ensembled multiple long short-term memory neural networks with different training times (i.e., LSTMes).

[0044] Optionally, when the single-mode ensemble model is a single-mode multilayer perceptron model (i.e., MLPe), the average value of the first predicted carbon emission data is obtained using the following formula: , Where y^` is the average value of the first predicted carbon emission data, N is the number of MLPs in the single-mode ensemble model, i=1 means accumulating from the first MLP, and y^MLP i The first predicted carbon emission data is output by the i-th multilayer sensor.

[0045] When the single-mode ensemble model is a single-mode recurrent neural network model (i.e., RNNe), the average value of the first predicted carbon emission data is obtained using the following formula: , Where y^` is the average value of the first predicted carbon emission data, N is the number of RNNs in the single-mode ensemble model, i=1 means accumulating from the first RNN, and y^RNN i This is the first predicted carbon emission data output by the i-th recurrent neural network.

[0046] When the single-mode ensemble model is a single-mode long short-term memory neural network model (i.e., LSTMe), the average value of the first predicted carbon emission data is obtained using the following formula: , Where y^` is the average value of the first predicted carbon emission data, N is the number of LSTMs in the single-mode ensemble model, i=1 means accumulating from the first LSTM, and y^LSTM i This is the first predicted carbon emission data output by the i-th long short-term memory neural network.

[0047] Alternatively, when the interleaved ensemble model is an interleaved multilayer perceptron model (i.e., MLPes), the average value of the first predicted carbon emission data is obtained using the following formula: , Where yˆ` is the average value of the first predicted carbon emission data, N is the number of multilayer sensors, yˆ is the first predicted carbon emission data obtained by the multilayer sensors, and x represents different years.

[0048] When the interleaved ensemble model is an interleaved recurrent neural network model (i.e., RNNes), the average value of the first predicted carbon emission data is obtained using the following formula: , Where yˆ` is the average value of the first predicted carbon emission data, N is the number of recurrent neural networks, yˆ is the first predicted carbon emission data obtained by the recurrent neural network, and x represents different years.

[0049] When the interleaved ensemble model is an interleaved long short-term memory neural network model (i.e., LSTMes), the average value of the first predicted carbon emission data is obtained using the following formula: , Where yˆ` is the average value of the first predicted carbon emission data, N is the number of long short-term memory neural networks, yˆ is the first predicted carbon emission data obtained by the long short-term memory neural network, and x represents different years.

[0050] Preferably, when the interleaved multilayer perceptron model (i.e., MLPes) is formed by integrating three multilayer perceptrons (i.e., MLPs) trained for three different training times (i.e., trained on four years of carbon dioxide emission data, five years of carbon dioxide emission data, and six years of carbon dioxide emission data, respectively), the average value of the first predicted carbon emission data is obtained using the following formula: , Where yˆ` is the average value of the first predicted carbon emission data, N is the number of multilayer sensors, and yˆ is the first predicted carbon emission data obtained by the multilayer sensors.

[0051] In this embodiment, the interleaved multilayer perceptron model (i.e., MLPes) is formed by integrating multiple multilayer perceptrons (i.e., MLPs) with different training times, the interleaved recurrent neural network model (i.e., RNNes) is formed by integrating multiple recurrent neural networks (i.e., RNNs) with different training times, and the interleaved long short-term memory neural network model (i.e., LSTMes) is formed by integrating multiple long short-term memory neural networks (i.e., LSTMs) with different training times. Each interleaved model can obtain the average value of the prediction results for different years, enabling this interleaved model to capture both short-term and long-term patterns.

[0052] In this embodiment, the average value of the first predicted carbon emission data obtained by the first integrated model, the second integrated model, or the third integrated model is equivalent to the second predicted carbon emission data that is close to the actual carbon emission data.

[0053] Furthermore, in step S304, the evaluation model includes actual carbon dioxide data and various data charts. After importing the second predicted carbon emission data obtained in step S303 into the evaluation model, data charts of the required style can be formed based on the actual carbon dioxide data and the second predicted carbon emission data. In the data charts, if the value point of the second predicted carbon emission data is adjacent to the value point of the actual carbon dioxide data, then the second predicted carbon emission data is determined to be accurate; if the value point of the second predicted carbon emission data is far from the value point of the actual carbon dioxide data, then the second predicted carbon emission data is determined to be inaccurate.

[0054] Application Example 1 like Figures 2 to 7 As shown, exemplarily, steps S200 and S300 in this embodiment are as follows: 1) Data Analysis Figure 2This diagram illustrates the correlations of CO2 emissions from 2010 to 2022. Each cell in the matrix represents the correlation coefficient between CO2 emissions in one year and those in another, ranging from 0.86 to 1.00. Most years show high correlation coefficients (above 0.9) with adjacent years, indicating that CO2 emission patterns tend to be consistent over time. For example, emissions in 2010 are highly correlated with emissions in 2011 and 2012, suggesting a continuous emission trend. As the years progress (e.g., comparing 2010 with 2022), the correlation decreases, represented by a cooler color in the matrix. This decline in correlation suggests that emissions may fluctuate more significantly over the long term, possibly due to changes in industrial activity and technological advancements affecting emissions. The correlation decreases slightly from 2020 to 2022 compared to earlier years, likely due to the impact of global events such as the COVID-19 pandemic, which temporarily disrupted industrial activity and transportation, leading to unusual emission patterns.

[0055] Figure 3 and Figure 4 It shows the CO2 emission trends and annual percentage changes for selected facilities from 2010 to 2022. Figure 3 The line graph shows the absolute CO2 emissions (in kg) of five facilities over time, identified by their facility IDs. For example, facility ID 100543.0 has a higher emission level, likely due to changes in operation or production. Facilities IDs 1003188.0 and 1003739.0 have relatively stable emission levels, indicating that their operation or production processes have not changed much over the years.

[0056] Figure 4 The line graph in the figure shows the percentage change in annual CO2 emissions for the same group of facilities. The annual changes highlight the volatility of emissions from each facility, with some years showing sharp increases or decreases.

[0057] For example, facility ID 100112.0 showed a significant decrease in emissions in some years (such as 2014) and a substantial increase in others. This may indicate that operational adjustments or regulatory impacts have affected emissions. Facility ID 1003739.0 also shows volatility, but its fluctuations are smaller compared to facility ID 100112.0, which may indicate that emissions management is more controllable.

[0058] 2) Carbon emission forecasting 2.1) Deep learning models: Figure 5This is a graph showing carbon emission predictions using multilayer perceptron, recurrent neural network, and long short-term memory network models. Comparing the different model predictions with actual data (lines marked with dots), it can be seen that the multilayer perceptron (lines marked with X) predicts emissions that remain almost constant across all years, significantly lower than the actual values.

[0059] This underestimation and lack of variation suggests that multilayer perceptrons (MLPs) may fail to effectively capture temporal trends or patterns in the data. MLPs lack the memory capacity of recurrent neural networks (RNNs) or long short-term memory networks (LSTMs), which may limit their ability to model sequence dependencies. The predictions of recurrent neural networks (RNNs) (lined with triangle markers) are closer to the overall trend of the true values ​​than those of MLPs. While not capturing exact values, they do show a downward trend in emissions from 2016 to 2022, with a significant drop around 2020. However, RNNs struggle to capture finer details and fluctuations, suggesting that the model partially captures sequence patterns but lacks precision, likely due to the vanishing gradient problem common in standard recurrent neural networks. The predictions of long short-term memory networks (LSTMs) deviate most significantly from the true values. Its predictions remain very low and almost constant, similar to multilayer perceptrons (MLPs), but with less variation across years. Overall, recurrent neural networks (RNNs) are the best performing deep learning model in capturing the general trend of CO2 emissions.

[0060] 2.2) Integration Model: Figures 6a to 6e To illustrate the prediction results of the ensemble models, a comparison is shown between emission data from a single-mode recurrent neural network model (RNNe (i.e., triangle markers)), a second ensemble model (MR (i.e., square markers)), and an interleaved multilayer perceptron model (i.e., MLPes (star markers)) and actual emission data (i.e., circle markers).

[0061] Each test case represents a different dataset sample, including actual emissions data from 2016 to 2022 (i.e., circled). It is worth noting that... Figure 6c The emissions in this case were significantly higher than in other cases, indicating the diversity of emission scenarios in the dataset. The interleaved multilayer perceptron model (i.e., MLPes (star markers)) showed some predictive accuracy, but generally lower than the emissions in all test cases. While the interleaved multilayer perceptron model (i.e., MLPes (star markers)) captured some trends, it lacked the precision required for a close match to the true emission data (i.e., circular markers), especially in... Figure 6b and Figure 6c In the meantime, it consistently falls below predictions. Interleaved multilayer perceptron models (i.e., MLPes (star-tagged)) typically struggle with continuous data, which can be explained by the fact that interleaved multilayer perceptron models (i.e., MLPes (star-tagged)) themselves do not capture temporal dependencies, resulting in lower accuracy.

[0062] Compared to interleaved multilayer perceptron models (i.e., MLPes (star markers)), single-mode recurrent neural network models (i.e., RNNe (triangle markers)) show better consistency with real emissions data (i.e., circular markers), capturing the overall trend of emissions in most test cases. Figure 6a , Figure 6b and Figure 6d In this study, the single-mode recurrent neural network model (i.e., RNNe (triangle-labeled)) follows the general trend of emissions and provides relatively close predictions, although some biases still exist. The single-mode recurrent neural network model (i.e., RNNe (triangle-labeled)) in... Figure 6c The study shows that it can handle larger emission values ​​better than interleaved multilayer perceptron models (i.e., MLPes (star-tagged)), but still remains low in some years. This improvement over interleaved multilayer perceptron models (i.e., MLPes (star-tagged)) is the ability to capture temporal patterns in single-mode recurrent neural network models (i.e., RNNe (triangle-tagged)), which is beneficial for emission data that may follow sequential patterns.

[0063] Among the ensemble models, the single-mode recurrent neural network model (i.e., RNNe (triangle markers)) performed best overall because it captured emission trends more accurately in most test cases. However, it still suffers from underestimation in high-emission scenarios. The second ensemble model (i.e., MR) provided stable but sometimes overly high predictions, while the interleaved multilayer perceptron model (i.e., MLPes (star markers)) performed poorly due to its inability to capture time dependencies.

[0064] Figure 7A comparative plot shows the predictive performance of various models using root mean square error (RMSE) as the evaluation metric. The Multilayer Perceptron (MLP) shows a moderate RMSE compared to other models, indicating it has an acceptable error level, but may not be the most accurate model due to its lack of time-dependent memory. The Recurrent Neural Network (RNN) has a lower RMSE than the MLP, indicating it effectively captures temporal patterns in the data. However, it is not the best performer, as some ensemble models achieve even lower RMSEs. The Long Short-Term Memory (LSTM) Neural Network has a significantly higher RMSE than both the MLP and RNN models.

[0065] The single-mode multilayer perceptron model (MLPe) and the interleaved multilayer perceptron model (MLPes) exhibit relatively low root mean square errors (RMSEs), indicating that ensembles improve predictive performance by reducing variability and bias compared to multilayer perceptrons (MLPs). The single-mode recurrent neural network model (RNNe) and the interleaved recurrent neural network model (RNNes) also demonstrate low RMSEs, with the interleaved recurrent neural network model (RNNes) performing slightly better. This highlights the effectiveness of recurrent neural networks (RNNs) in capturing time-series patterns, with lower prediction errors compared to recurrent neural networks (RNNs). The single-mode long short-term memory neural network model (LSTMe) and the interleaved long short-term memory neural network model (LSTMes) exhibit very high RMSEs, similar to long short-term memory neural networks (LSTMs), further suggesting that long short-term memory neural networks (LSTMs) may not be suitable for the observations on this dataset.

[0066] The first ensemble model (AE), which combines multiple models, has one of the highest root mean square error values. This may indicate that combining poorly performing models, such as Long Short-Term Memory (LSTM) neural networks, could negatively impact the overall accuracy of the ensemble. The second ensemble model (MR) has the lowest root mean square error among all models, suggesting that this particular combination effectively balances stability and time sensitivity to provide the most accurate predictions.

[0067] Overall, by Figure 7To date, the second ensemble model (MR) outperforms complex models such as Long Short-Term Memory Neural Networks (LSTMs), and even surpasses single-mode LSTM models (LSTMe) and interleaved LSTM models (LSTMes). The second ensemble model (MR) successfully integrates Multilayer Perceptrons (MLPs) and Recurrent Neural Networks (RNNs) to achieve a balance between accuracy and model stability, demonstrating that simpler and well-tuned models can sometimes outperform complex architectures in CO2 emission prediction tasks.

[0068] Example 2 like Figure 8 As shown, this embodiment provides a prediction device for implementing the annual carbon emission prediction method described in Embodiment 1 above, including a data processing module 1, a data analysis module 2, and a carbon emission prediction model 3, wherein: The data processing module 1 is used to process the obtained dataset according to preset conditions to obtain a data subset that includes at least carbon emission information and different years, and to execute the specific steps of S100 in Embodiment 1.

[0069] Data analysis module 2 is used to analyze a subset of data, form a visual representation of carbon emission information for different years, and form an evaluation matrix to estimate the relationship between carbon emission information and different years, in order to perform the specific steps of S200 in Example 1.

[0070] Carbon emission prediction model 3 is used to construct a carbon emission prediction model that includes a deep learning model and an ensemble model. When a subset of data is input into the carbon emission prediction model, a carbon emission trend prediction result that is close to the real data is obtained, which is used to execute the specific steps of S300 in Example 1.

[0071] Furthermore, in carbon emission prediction module 3, there are a construction module, a deep learning model, an ensemble model, and an evaluation model, wherein: The building module constructs a carbon emission prediction model that includes deep learning and ensemble models, which is used to execute the specific steps of S301 in Example 1. The deep learning model includes multiple learning models, and a subset of data is input into each learning model to obtain multiple first predicted carbon emission data, which are used to execute the specific steps of S302 in Example 1. The integrated model receives at least one first predicted carbon emission data as input to derive a second predicted carbon emission data that approximates the actual carbon emission data, for performing the specific steps of S303 in Example 1. The evaluation model is used to evaluate the second predicted carbon emission data. If the second predicted carbon emission data is determined to meet the comparison conditions, the second predicted carbon emission data is determined to be accurate, and the specific steps of S304 in Example 1 are executed.

[0072] Example 3 like Figure 9 As shown, this embodiment also provides an electronic device, including: a communication interface 5, which can interact with other devices such as network devices; The processor 6 is connected to the communication interface 5 to enable information exchange with other devices and to execute the steps of the aforementioned annual carbon emission prediction method when running a computer program. The computer program is stored in the memory 4.

[0073] Of course, in practical applications, the various components in an electronic device are coupled together through bus system 7. It can be understood that bus system 7 is used to achieve communication and connection between these components. In addition to the data bus, bus system 7 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in... Figure 9 The general will label all buses as Bus System 7.

[0074] In this embodiment, memory 4 is used to store various types of data to support the operation of the electronic device. Examples of such data include any computer program used to operate on the electronic device.

[0075] It is understood that memory 4 can be volatile memory or non-volatile memory, or both. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), ferromagnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); magnetic surface memory can be disk storage or magnetic tape storage. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Synchronous Static Random Access Memory (SSRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDRSDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), SyncLink Dynamic Random Access Memory (SLDRAM), and Direct Rambus Random Access Memory (DRRAM).The memory 4 described in the embodiments of this application is intended to include, but is not limited to, these and any other suitable types of memory.

[0076] The methods disclosed in the embodiments of this application can be applied to processor 6, or implemented by processor 6. Processor 6 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in processor 6 or by instructions in software form. The processor 6 may be a general-purpose processor, DSP, or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 6 can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the methods disclosed in the embodiments of this application can be directly manifested as being executed by a hardware code processor, or being executed by a combination of hardware and software modules in the code processor. The software modules may be located in a storage medium, which is located in memory 4. Processor 6 reads the program in memory 4 and combines it with its hardware to complete the steps of the aforementioned method.

[0077] When processor 6 executes the program, it implements the corresponding process of the annual carbon emission prediction method described in Embodiment 1 above. For the sake of brevity, it will not be described in detail here.

[0078] Example 4 This application also provides a storage medium, namely a computer storage medium, specifically a computer-readable storage medium, such as a memory that stores a computer program, which can be executed by a processor to complete the steps described in the aforementioned method. The computer-readable storage medium can be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, Flash Memory, magnetic surface memory, optical disc, CD-ROM, etc.

[0079] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the annual carbon emission prediction method described in Embodiment 1 above. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0080] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device (which may be a personal computer, server, network device, etc.) to execute all or part of the annual carbon emission prediction method described in Embodiment 1 above. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0081] The above description is merely a preferred embodiment of the present invention and is illustrative rather than restrictive. Those skilled in the art will understand that many changes, modifications, and even equivalents can be made within the spirit and scope defined by the claims of the present invention, all of which will fall within the protection scope of the present invention.

Claims

1. A method for predicting annual carbon emissions, characterized in that, Includes the following steps: The dataset is processed according to preset conditions to obtain a subset of data that includes at least carbon emission information and data from different years. Analyze the data subset to form a visual diagram of the carbon emission information for different years, and form an evaluation matrix for estimating the relationship between the carbon emission information and the different years; A carbon emission prediction model is constructed that includes a deep learning model and an ensemble model. When the data subset is input into the carbon emission prediction model, a carbon emission trend prediction result that closely approximates the real data is obtained.

2. The annual carbon emission forecasting method according to claim 1, characterized in that, The process of processing the dataset according to preset conditions to obtain a data subset that includes at least carbon emission information and data from different years includes the following steps: The obtained dataset is filtered according to the first preset condition, and data that is related to the first preset condition is deleted to form the first dataset; The first dataset is sorted according to the second preset condition to form a sequence corresponding to the second preset condition; The sequence is filtered according to a third preset condition to form a subset of data that does not contain the third preset condition.

3. The annual carbon emission forecasting method according to claim 1, characterized in that, The process of constructing a carbon emission prediction model that includes a deep learning model and an ensemble model, and inputting the data subset into the carbon emission prediction model to obtain a carbon emission trend prediction result that approximates the real data, includes the following steps: Construct the carbon emission prediction model that includes the deep learning model and the ensemble model; The data subsets are respectively input into the deep learning model to obtain multiple first predicted carbon emission data; At least one of the first predicted carbon emission data is input into the integrated model to obtain a second predicted carbon emission data that is close to the actual carbon emission data. The second predicted carbon emission data is evaluated to determine its accuracy.

4. The annual carbon emission forecasting method according to claim 3, characterized in that, The deep learning model includes a multilayer perceptron, a recurrent neural network, and a long short-term memory neural network, wherein: After the data subset is input into the multilayer sensor, the first predicted carbon emission data is obtained using the following formula: yˆ=f(WL···f(Wif(Wix+bi)+bi)···+bL), Where yˆ is the first predicted carbon emission data, f is the activation function, Wi and bi are the weights and biases of the i-th layer, i is a number, and L is the layer number; After the data subset is input into the recurrent neural network, the first predicted carbon emission data is obtained sequentially using the following formula. ht=f(Wxxt+Whht-1+b), Among them, h t Let be the hidden state vector at time t, f be the activation function, Wx, Wh, and b be the learning functions, xt be the input vector at time t, and h be the learning vector. t -1 represents the hidden state at the previous time step, and b is the bias vector; yˆt=g(Wyht+c), Where yˆt represents the first predicted carbon emission data at time t, g is the activation function, Wy and c are the learning functions, and h t Let be the hidden state vector at time t; After the data subset is input into the long short-term memory neural network, the first predicted carbon emission data is obtained sequentially using the following formula. ft = σ(Wf[ht-1, xt] + bf), Where ft is the forget gate vector, σ is the Sigmoid activation function, Wf is the weight matrix of the forget gate, and [ht] [1, xt] is the concatenation vector, and bf is the bias vector of the forget gate; it = σ(Wi[ht-1, xt] + bi), Where it is the input gate vector, σ is the Sigmoid activation function, Wi is the input gate weight matrix, [ht-1, xt] is the concatenation vector, and bi is the input gate bias vector; ct=ft⊙ct-1+it⊙tanh(Wc[ht 1,xt]+bc), Where ct is the cell state at the current time step, ft is the forgetting gate vector, ct-1 is the cell state at the previous time step, it is the input gate vector, and tanh(Wc[ht-1, xt]+bc) is the candidate cell state, where: Wc is the weight matrix, [ht-1, xt] is the vector concatenation of the previous hidden state ht-1 and the current input xt, bc is the bias term, tanh is the hyperbolic tangent activation function, and ⊙ is the Hadamard element-wise multiplication operator. ot = σ(Wo[ht-1, xt] + bo), Where ot is the output gate vector, σ is the Sigmoid activation function, Wo is the output gate weight matrix, [ht-1, xt] is the vector concatenation, and bo is the output gate bias vector; ht = ot⊙tanh(ct), Where ht is the hidden state at the current time step, ot is the output gate vector, ⊙ is the Hadamard element-wise multiplication operator, tanh is the element-wise function, and ct is the vector.

5. The annual carbon emission forecasting method according to claim 4, characterized in that, The integrated model is divided into a first integrated model, a second integrated model and a third integrated model. The first integrated model is formed by integrating the multilayer perceptron, the recurrent neural network and the long short-term memory neural network, and is used to obtain the average value of the first predicted carbon emission data. The second integrated model is formed by integrating two of the multilayer perceptron, the recurrent neural network, and the long short-term memory neural network, and is used to obtain the average value of the first predicted carbon emission data; The third integrated model is formed by integrating multiple multilayer perceptrons, or multiple recurrent neural networks, or multiple long short-term memory neural networks, and is used to obtain the average value of the first predicted carbon emission data.

6. The annual carbon emission forecasting method according to claim 5, characterized in that, The third integration model is either a single-mode integration model or an interleaved integration model, wherein: The single-mode ensemble model is formed by integrating multiple multilayer perceptrons with different training parameters, or multiple recurrent neural networks with different training parameters, or multiple long short-term memory neural networks with different training parameters. The interleaved ensemble model is formed by integrating multiple multilayer perceptrons with different training times, or multiple recurrent neural networks with different training times, or multiple long short-term memory neural networks with different training times.

7. A prediction device, characterized in that, For implementing the annual carbon emission forecasting method according to any one of claims 1 to 6, comprising: The data processing module is used to process the obtained dataset according to preset conditions to obtain a data subset that includes at least carbon emission information and data from different years. The data analysis module is used to analyze the data subset, form a visual representation of the carbon emission information for different years, and form an evaluation matrix for estimating the relationship between the carbon emission information and the different years. A carbon emission prediction model is used to construct a carbon emission prediction model that includes a deep learning model and an ensemble model. When the data subset is input into the carbon emission prediction model, a carbon emission trend prediction result that closely approximates the real data is obtained.

8. The prediction device according to claim 7, characterized in that, The carbon emission prediction module includes a construction module, a deep learning model, an ensemble model, and an evaluation model, wherein: The building module is used to construct a carbon emission prediction model that includes deep learning and ensemble models; The deep learning model is used to derive multiple first predicted carbon emission data based on the data subset; The integrated model receives at least one of the first predicted carbon emission data as input to derive a second predicted carbon emission data that closely approximates the actual carbon emission data. The evaluation model is used to evaluate the second predicted carbon emission data and determine the accuracy of the second predicted carbon emission data.

9. A medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the annual carbon emission prediction method according to any one of claims 1 to 6.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the annual carbon emission prediction method according to any one of claims 1 to 6.

Images