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

Abstract

Greenhouse gas emissions, as one of the primary contributors to global warming, present an urgent environmental challenge that requires attention. Accurate prediction of carbon dioxide (CO2) emissions from the industrial sector is crucial for the development of low-carbon industries. However, existing time series models often suffer from severe overfitting when data volume is insufficient. In this paper, we propose a carbon emission prediction method based on meta-learning and differential long- and short-term memory (MDL) to address this issue. Specifically, MDL leverages Long Short-Term Memory (LSTM) to capture long-term dependencies in time series data and employs a meta-learning framework to transfer knowledge from multiple source task datasets for initializing the carbon emission prediction model for the target task. Additionally, the combination of differential LSTM and the meta-learning framework reduces the dependency of the differential long- and short-term memory network on data volume. The smoothed difference method, included in this approach, mitigates the randomness of carbon emission sequences, consequently benefiting the fit of the LSTM model to the data. To evaluate the effectiveness of our proposed method, we validate it using carbon emission datasets from 30 provinces in China and the industrial sector in Xinjiang. The results show that the average absolute error (MAE), Coefficient of Determination (R2) and root mean square error (RMSE) of the method have been reduced by 61.8% and 63.8% on average compared with the current mainstream algorithms. The method provides an efficient and accurate solution to the task of industrial carbon emission prediction, and helps environmental policy makers to formulate environmental policies and energy consumption plans.