关键词: 多尺度Transformer, 自适应频率解耦, 原油价格预测, 趋势融合, 低碳经济

Abstract: Against the backdrop of global low-carbon transition and energy restructuring,crude oil price forecasting has become not only a key topic in energy market analysis but also an essential reference for policy-making and investment decisions.How-ever,crude oil price series often exhibit strong nonlinearity and pronounced non-stationarity.Existing methods still face limitations in feature extraction and temporal modeling:on the one hand,the mining of multi-scale features is often insufficient,leading to biased characterization of short-term fluctuations and long-term trend evolution;on the other hand,the integration of short- and long-term information is frequently mishandled,making it difficult to balance predictive accuracy with trend stability.To address these challenges,this paper proposes a multi-frequency decoupled dual-branch Transformer model(MFD-DBV-Transformer) for Brent crude oil price forecasting.The method firstly employs adaptive empirical mode decomposition(AEMD) to decompose the crude oil price series into multiple intrinsic mode functions(IMFs).This distinguishes high-frequency short-term components from low-frequency long-term trends.An adaptive frequency decoupling module(AFDM) is then designed to construct dual-branch feature representations,separately capturing short-term volatility patterns and long-term trend features.A trend fusion module is further introduced,where cross-attention is used to achieve adaptive modulation of short-term predictions with long-term trend information.In addition,a temperature-regulated adaptive masking mechanism is incorporated to prevent overfitting in long-term trend modeling and to enhance the model’s generalization ability in volatile market environments.Experimental results demonstrate that the proposed MFD-DBV-Transformer achieves superior performance in capturing complex time-frequency characteristics of crude oil prices,significantly outperforming traditional LSTM and several mainstream deep learning models.The model not only improves forecasting accuracy but also demonstrates stronger stability and adaptability in trend tracking.The proposed approach provides policymakers and energy investors with an efficient and reliable forecasting and decision-support tool for coping with crude oil market volatility,while offering new insights and methodologies for modeling complex non-stationary time series.

Key words: Multi-scale transformer, Adaptive frequency decoupling, Crude oil price forecasting, Trend fusion, Low-carbon economy