MDL: Industrial carbon emission prediction method based on meta-learning and diff long short-term memory networks

Area of Science:

• Environmental Science
• Data Science
• Machine Learning

Background:

• Greenhouse gas emissions, particularly carbon dioxide (CO2), are major drivers of global warming.
• Accurate CO2 emission prediction in the industrial sector is essential for developing effective low-carbon strategies.
• Current time series models struggle with overfitting when dealing with insufficient data for industrial emissions.

Purpose of the Study:

• To propose an advanced carbon emission prediction method that overcomes data limitations.
• To improve the accuracy and efficiency of industrial CO2 emission forecasting.
• To support environmental policy formulation and energy planning.

Main Methods:

• Developed a novel method combining meta-learning with differential Long Short-Term Memory (LSTM) networks (MDL).
• Utilized LSTM to capture long-term dependencies in time series data.
• Employed a meta-learning framework for knowledge transfer from diverse datasets to initialize the target prediction model.
• Incorporated a smoothed difference method to stabilize carbon emission sequences and improve model fitting.

Main Results:

• The MDL method demonstrated significant improvements in prediction accuracy.
• Achieved average reductions of 61.8% in Mean Absolute Error (MAE) and 63.8% in Root Mean Square Error (RMSE) compared to mainstream algorithms.
• Validated using carbon emission datasets from 30 Chinese provinces and the Xinjiang industrial sector.

Conclusions:

• The proposed MDL method offers an efficient and accurate solution for industrial carbon emission prediction.
• The approach effectively mitigates issues of overfitting and data dependency in time series modeling.
• This method provides valuable insights for policymakers in developing environmental and energy consumption strategies.