An Ensemble Prediction System Based on Artificial Neural Networks and Deep Learning Methods for Deterministic and Probabilistic Carbon Price Forecasting

Abstract

Carbon price prediction is important for decreasing greenhouse gas emissions and coping with climate change. At present, a variety of models are widely used to predict irregular, nonlinear, and nonstationary carbon price series. However, these models ignore the importance of feature extraction and the inherent defects of using a single model; thus, accurate and stable prediction of carbon prices by relevant industry practitioners and the government is still a huge challenge. This research proposes an ensemble prediction system (EPS) that includes improved data feature extraction technology, three prediction submodels (GBiLSTM, CNN, and ELM), and a multiobjective optimization algorithm weighting strategy. At the same time, based on the best fitting distribution of the prediction error of the EPS, the carbon price prediction interval is constructed as a way to explore its uncertainty. More specifically, EPS integrates the advantages of various submodels and provides more accurate point prediction results; the distribution function based on point prediction error is used to establish the prediction interval of carbon prices and to mine and analyze the volatility characteristics of carbon prices. Numerical simulation of the historical data available for three carbon price markets is also conducted. The experimental results show that the ensemble prediction system can provide more effective and stable carbon price forecasting information and that it can provide valuable suggestions that enterprise managers and governments can use to improve the carbon price market.