A novel deep learning carbon price short-term prediction model with dual-stage attention mechanism

Abstract

• First utilization of a complete deep learning framework to make short-term predictions of carbon price time series data. • Replace the feature engineering work of traditional machine learning with dual-stage attention mechanisms. • Decompose the price series using CEEMDAN and reduce the dimensionality of the input series with PCA . • The influence of model hyperparameters on prediction results is discussed. Carbon price prediction can help participants keep abreast of carbon market dynamics and develop trading strategies. It is challenging for statistical models to accurately capture the nonlinear characteristics of the carbon pricing, and machine learning methods need sophisticated artificial feature engineering. To successfully address these drawbacks, our research suggests a carbon price forecasting model built on a deep learning architecture. Complete Ensemble Empirical Mode Decomposition with Adaptive Noise decomposes historical price to obtain Intrinsic Mode Function and Principal Component Analysis reduces the dimensionality of each influential factor. Dual-Stage Attention-Based Recurrent Neural Network, a Seq2Seq model, made up of an encoder with feature attention and a decoder with temporal attention, is employed to predicted price of the Hubei Carbon Emissions Allowance. The dual-attention mechanism enables preprocessing to be done adaptively and more effectively than manual processing. As shown by statistical analysis and grey correlation analysis, Hubei Carbon Emissions Allowance has a high autocorrelation, and the carbon market, energy and industry, economy, and environment have high to low correlations on it. The accuracy metrics of this framework, Mean Absolute Error = 0.75, Mean Absolute Percentage Error = 1.59 and Root Mean Squared Error = 1.28, are lower than compared models.