LightGBM-BES-BiLSTM Carbon Price Prediction Based on Environmental Impact Factors

Abstract

A carbon trading price fusion prediction model is proposed to capture the non-linear, non-stationary, multi-frequency, and other irregular characteristics of carbon price data, as well as the temporal periodicity of environmental factors. Firstly, an adaptive Symmetric geometric mode decomposition method is introduced to address the irregularities in carbon trading prices, including nonlinearity, non-stationarity, and multi-frequency. Bubble entropy is employed to extract global features in the frequency and time domains of carbon price data. Secondly, to handle the nonlinearity, temporal periodicity, and noise in environmental influencing factors, a mapping function between the frequency components of carbon price data and environmental influencing factors is established using LightGBM (Light gradient boosting machine) with a regularization term, enabling enhanced fusion of carbon price data features. Thirdly, a Bald Eagle Search-optimized Bi-directional long short-term memory (BiLSTM) model is proposed for predicting carbon prices with different cycle and frequency components. Finally, experimental results demonstrate the superior performance of the proposed fusion prediction model over other models.