CN-122022131-A - Carbon emission prediction method and system based on multi-model fusion

CN122022131ACN 122022131 ACN122022131 ACN 122022131ACN-122022131-A

Abstract

The invention discloses a carbon emission prediction method and a system based on multi-model fusion, wherein the method comprises the steps of constructing a multi-dimensional data set containing time sequence characteristics, cross characteristics and industry characteristics through preprocessed power equipment production enterprise data; the method comprises the steps of constructing a carbon emission prediction model based on a multidimensional data set, wherein the carbon emission prediction model comprises three sub-model linear regression models, a random forest model and an improved long-short-term memory network model which are arranged independently of each other, the improved long-short-term memory network model integrates residual connection and attention mechanisms, dynamically calculating weights of the linear regression model, the random forest model and the improved long-short-term memory network model based on a variance-covariance matrix of a historical prediction error of the carbon emission prediction model, obtaining a carbon emission comprehensive prediction value based on the weights of the three sub-models, and generating an emission reduction strategy based on the carbon emission comprehensive prediction value. The method and the device can improve the prediction precision and generalization capability of the carbon emission prediction model.

Inventors

MA LINGJIE
TANG CHENGJUN
JI LING
LI RUI
YU FAN

Assignees

南京国电南自电网自动化有限公司

Dates

Publication Date: 20260512
Application Date: 20260107

Claims (10)

1. A carbon emission prediction method based on multi-model fusion is characterized by comprising the following steps: Acquiring preprocessed data of an electric equipment production enterprise; Constructing a multidimensional data set containing time sequence characteristics, crossing characteristics and industry characteristics based on the preprocessed power equipment production enterprise data; The carbon emission prediction model is constructed based on the multidimensional dataset, wherein the carbon emission prediction model comprises a linear regression model, a random forest model and an improved long-short-term memory network model which are arranged independently of each other, and the improved long-short-term memory network model integrates residual error connection and attention mechanisms; dynamically calculating weights of a linear regression model, a random forest model and an improved long-term and short-term memory network model based on a variance-covariance matrix of a historical prediction error of the carbon emission prediction model; obtaining a comprehensive carbon emission predicted value based on the weights of the linear regression model, the random forest model and the improved long-term and short-term memory network model; And generating an emission reduction strategy based on the comprehensive carbon emission predicted value.
2. The carbon emission prediction method based on multimodal fusion as defined in claim 1, wherein the electrical equipment production enterprise data includes historical carbon emission data, energy consumption data, equipment operating parameters, and environmental data; the preprocessing comprises missing value processing and abnormal value detection, wherein the missing value processing comprises filling missing data by adopting a cubic spline interpolation method or a K nearest neighbor algorithm, and the abnormal value processing comprises identifying and correcting abnormal data points by combining a 3 sigma criterion and a device operation parameter threshold.
3. The carbon emission prediction method based on multi-model fusion according to claim 1, wherein the time sequence features comprise statistics of average daily energy consumption, standard deviation of energy consumption and gradient of energy consumption in a specific past day through a sliding window, and the average daily energy consumption is calculated according to the following formula: , In the formula, For calculating the number of days of daily average energy consumption, the value range is 1- , As a total number of days, Is the first The energy consumption of the day is high, The energy consumption standard deviation calculation formula is as follows: , In the formula, Is that The daily average energy consumption of the day is that, The energy consumption slope calculation formula is as follows: , In the formula, For total days Is a median of (a).
4. The carbon emission prediction method based on multimodal fusion as defined in claim 1, wherein the cross features include a first cross feature and a second cross feature, and the calculation formulas are as follows: , , In the formula, As a first of the cross-over features, As a second feature of the cross-over, For the operating temperature of the device, For the purpose of energy consumption of the device, In order to produce the order quantity, The number of times of starting and stopping the equipment.
5. The multimodal fusion-based carbon emission prediction method as defined in claim 1, wherein the industry characteristics include provincial grid carbon emission factors Regional industrial electricity price fluctuation index Policy adjustment coefficients 。
6. The carbon emission prediction method based on multimodal fusion as defined in claim 1, wherein the expression of the residual connection is as follows: , In the formula, In order to be able to take the moment of time, Is connected by residual errors The hidden state of the moment of time, In order to be in the hidden state at the last moment, Is that The input characteristics of the time of day, The expression of the attention mechanism is as follows: , , In the formula, Is that The attention weight of the moment in time, Is that The attention score of the moment in time, For the moment of time The value range is 1-B, B is the time sequence length, In order for the parameters to be trainable, To be a transpose of a trainable parameter vector, Is index of Is a score of attention of (a).
7. The method for predicting carbon emissions based on multimodal fusion of claim 1, wherein dynamically calculating weights of a linear regression model, a random forest model and an improved long-short term memory network model based on a variance-covariance matrix of a historical prediction error of the carbon emissions prediction model comprises dynamically adjusting weights of the linear regression model, the random forest model and the improved long-short term memory network model at specific intervals, wherein a calculation formula of the variance-covariance matrix is as follows: , In the formula, As the variance-covariance matrix, For calculating the number of days of the variance-covariance matrix, the value range is 1- , As a total number of days, Model number of carbon emission prediction The prediction error vector generated by the prediction result of the day, Is the error mean, and T is the transposed symbol.
8. The carbon emission prediction method based on multi-model fusion according to claim 7, wherein the weight calculation formulas of the linear regression model, the random forest model and the improved long-term and short-term memory network model are as follows: , , In the formula, The values of the sequence numbers of the linear regression model, the random forest model and the improved long-short-term memory network model are 1-3, Is the first The weight of the individual model is determined, Is a combined weight vector of linear regression model weight, random forest model weight and improved long-term and short-term memory network model weight, The expression of (2) is as follows: 。
9. The carbon emission prediction method based on multimodal fusion according to claim 8, wherein the comprehensive carbon emission prediction value is calculated as follows: , In the formula, Is a comprehensive predicted value of the carbon emission, The carbon emission predicted values of the linear regression model, the random forest model and the improved long-term and short-term memory network model are respectively obtained.
10. A carbon emission prediction system based on multimodal fusion, comprising: The data preprocessing module is used for acquiring preprocessed data of the power equipment manufacturing enterprises; the data set construction module is used for constructing a multi-dimensional data set containing time sequence characteristics, cross characteristics and industry characteristics based on the preprocessed data of the power equipment manufacturing enterprise; The model construction module is used for constructing a carbon emission prediction model based on the multidimensional dataset, wherein the carbon emission prediction model comprises a linear regression model, a random forest model and an improved long-short-term memory network model which are arranged independently of each other, and the improved long-short-term memory network model integrates residual error connection and attention mechanisms; The dynamic fusion module is used for dynamically calculating weights of a linear regression model, a random forest model and an improved long-term and short-term memory network model based on a variance-covariance matrix of a historical prediction error of the carbon emission prediction model; The comprehensive prediction module is used for obtaining a carbon emission comprehensive prediction value based on the weights of the linear regression model, the random forest model and the improved long-term and short-term memory network model; and the strategy generation module is used for generating an emission reduction strategy based on the comprehensive carbon emission predicted value.

Description

Carbon emission prediction method and system based on multi-model fusion Technical Field The invention relates to the technical field of carbon emission prediction, in particular to a carbon emission prediction method and system based on multi-model fusion. Background The power plant production industry is an energy intensive industry with carbon emissions that account for a significant proportion of the industry. With the advancement of dual carbon targets, electrical equipment manufacturers face stringent carbon emission constraints. However, the existing carbon emission prediction methods have the following drawbacks: 1) In the traditional linear regression model, the linear relation among variables is assumed, but in actual production, the energy consumption, the yield, the temperature and other factors often show nonlinear relations, so that the prediction error is larger. 2) The random forest model can process nonlinear relation, but ignores time sequence characteristics of data, is poor in long sequence prediction and is easy to be interfered by noise data. 3) The standard LSTM model relieves the gradient problem through a gating mechanism, but the phenomenon of forgetting information can still occur when an ultra-long sequence is processed, and the training time is long. 4) And the multiple models are simply and evenly fused, namely, the dynamic weight distribution of the models is not considered, so that the stability of a prediction result is insufficient, and the prediction result cannot adapt to the change of working conditions. Disclosure of Invention The invention aims to provide a carbon emission prediction method and a system based on multi-model fusion, wherein a linear regression model, a random forest model and an improved long-term and short-term memory network model are built by pre-building a data set containing multi-dimensional characteristics, so that a carbon emission prediction model is built, the weight of the carbon emission prediction model is calculated and updated, a comprehensive carbon emission prediction value is calculated, and the prediction precision and generalization capability of the carbon emission prediction model are improved. The invention is realized by the following technical scheme. In a first aspect, the present invention provides a carbon emission prediction method based on multimodal fusion, comprising: Acquiring preprocessed data of an electric equipment production enterprise; Constructing a multidimensional data set containing time sequence characteristics, crossing characteristics and industry characteristics based on the preprocessed power equipment production enterprise data; The carbon emission prediction model is constructed based on the multidimensional dataset, wherein the carbon emission prediction model comprises a linear regression model, a random forest model and an improved long-short-term memory network model which are arranged independently of each other, and the improved long-short-term memory network model integrates residual error connection and attention mechanisms; dynamically calculating weights of a linear regression model, a random forest model and an improved long-term and short-term memory network model based on a variance-covariance matrix of a historical prediction error of the carbon emission prediction model; obtaining a comprehensive carbon emission predicted value based on the weights of the linear regression model, the random forest model and the improved long-term and short-term memory network model; And generating an emission reduction strategy based on the comprehensive carbon emission predicted value. Optionally, the power equipment manufacturing enterprise data comprises historical carbon emission data, energy consumption data, equipment operation parameters and environmental data; the preprocessing comprises missing value processing and abnormal value detection, wherein the missing value processing comprises filling missing data by adopting a cubic spline interpolation method or a K nearest neighbor algorithm to ensure the continuity of a time sequence, and the abnormal value processing comprises identifying and correcting abnormal data points by combining a 3 sigma criterion and a device operation parameter threshold. Optionally, the time sequence feature comprises statistics of average daily energy consumption, standard deviation of energy consumption and slope of energy consumption in the past specific days through a sliding window, wherein the average daily energy consumption is calculated according to the following formula: , In the formula, For calculating the number of days of daily average energy consumption, the value range is 1-,As a total number of days,Is the firstDay energy consumption. The energy consumption standard deviation calculation formula is as follows: , In the formula, Is thatThe daily average energy consumption of the day is that, The energy consumption slope calculation formula is as follows: