二次分解策略的深度学习碳交易价格预测

随着我国碳排放权交易市场的逐渐完善，碳交易价格的准确预测将有助于构建更加稳定的市场环境，极大减少参与者的风险。针对当前碳交易价格预测难度大及现有的二次分解-集合策略不完善等问题，该文提出一种基于变分模态分解(variational mode decomposition, VMD)和经验小波变化(empirical wavelet transform, EWT)的二次分解预测策略，其中分别采用中心频率(central frequency, CF)和Lempel-Ziv复杂度计算作为分解层数的确定依据，样本熵(sample entropy, SE)作为第二次分解输入序列的重构依据，使用长短期记忆网络(long short-term memory, LSTM)和时序卷积网络(temporal convolutional network, TCN)作为预测模型，并结合海洋捕食者算法(marine predator algorithm, MPA)对模型进行参数优化。实验结果表明，V-LSTM-E-LSTM模型和V-TCN-E-TCN模型不仅在湖北碳交易价格的短期和长期预测中获得了最好的效果，而且在其它四个区域碳排放权交易市场也获得了较高的精度。但对于成立时间较短的全国碳排放权交易市场，V-TCN-E-TCN模型在短期预测中表现更佳，长期预测中效果更好的是V-TCN-E-LSTM模型。

Deep Learning Carbon Trading Price Forecasting With Secondary Decomposition Strategy

With the gradual improvement of the national carbon emission trading market, the accurate forecasting of the carbon emission trading price will help to establish a more stable market environment and greatly reduce the risk of participants. To address the current problems such as the difficulty of carbon trading price forecasting and the imperfection of the existing secondary decomposition-aggregation strategy, this paper proposed a new forecasting strategy. This strategy was based on variational mode decomposition (VMD) and empirical wavelet transform (EWT). In addition, central frequency (CF) and Lempel-Ziv complexity calculations were used in this strategy as the basis for determining the number of decomposition levels. Meanwhile, sample entropy (SE) was used as the basis for data reconstruction. And the reconstructed new data was served as input data for secondary decomposition. Then, this strategy used long short-term memory (LSTM) and temporal convolutional network (TCN) as forecasting models. And it was combined with marine predator algorithm (MPA) in the model for parameter optimization. The experimental results show that the V-LSTM-E-LSTM model and the V-TCN-E-TCN model not only achieve the best results in the short-term and long-term forecasting of the carbon trading price in Hubei, but also achieve high accuracy in the other four regional carbon emission allowance trading markets. However, in the national carbon emissions trading market with a relatively short establishment time, the V-TCN-E-TCN model performs better in short-term forecasting, and the V-TCN-E-LSTM model is more effective in long-term forecasting.