统计与数据科学学院知识库

随着全球气候变化问题日益严重，碳排放权交易作为一种有效的碳减排手段在全球范围内得到广泛实施。中国作为世界上最大的碳排放权交易市场之一，其碳价格走势对于国内外的碳减排和经济发展具有重要影响。实现碳排放权交易价格的准确预测，能够帮助政府在“双碳”目标计划下进行宏观政策调控，能够帮助企业制定更精准的生产计划，所以建立预测精度较高的碳价格预测模型具有重要的现实意义。本文在现有中国碳价格预测领域研究成果的基础上，以北京、湖北、深圳三个碳交易所碳价格预测为例，主要研究内容如下：

（1）本文提出一种同步多分解方法的预测框架，通过与单一分解方法的预测结果做对比验证同步多分解思路的可行性。首先将历史碳价格序列同时做EMD分解、OVMD分解和SSA分解，计算每个分解子序列的LZ复杂度值并根据肘部法则进行聚类重构，再利用Ada-Elman模型对重构序列进行预测，最后BP做非线性集成。

（2）本文提出一种基于多数据信息的多元分解碳价格预测模型。首先根据现有的中国碳价格影响因素研究文献筛选有关影响因素，与历史碳价格共同输入做多元经验模态分解（MEMD），得到分解子序列相同的数据集，计算每个子序列的多尺度散步熵值（MDE）重构为高频、低频，针对不同频率做Lasso筛选重要的影响因素进入预测模型，通过对比验证该模型的优越性。

（3）本文提出了一种加入百度搜索信息且融合多元分解、多分解方法的预测模型。百度作为中国受众群体最大的搜索引擎，其搜索指数能够相当程度上反应碳价格的受关注程度。首先通过碳价格的相关影响词选择百度搜索指数，与查阅文献选择的其他市场的影响因素做GBDT模型进行重要性排序；其次，被挑选的因素和历史碳价格共同输入多元经验模态分解和多元变分模态分解（MEMD/MVMD）中进行分解，将分解序列重构为高频、低频、趋势项；接着引入Pearson相关系数、主成分分析（PCA）、格兰杰因果检验（Granger Causality）对序列进一步做过滤筛选；最终过滤得到的序列输入卷积神经网络-双向长短时记忆神经网络预测模型（CNN-BiLSTM）进行预测。通过与其他模型进行对比验证该模型的优越性。

根据本文所作工作可以得出结论：（1）基于市场影响因素和百度搜索指数的碳价格预测是可靠有效的；（2）同步多分解方法能够提高碳价格预测模型的预测效果 ；（3）当基于分解重构框架的碳价格预测模型加入了外生因素输入时，需要对外生因素序列做二次筛选，否则会影响模型预测效果。

With the increasingly serious problem of global climate change, carbon emission trading as an effective means of carbon emission reduction has been widely implemented in the world. As one of the largest carbon emission trading markets in the world, China's carbon price trend has an important impact on carbon emission reduction and economic development both domestically and internationally. Accurate prediction of carbon emission trading price can help the government to carry out macro-policy regulation under the "dual carbon" target plan and help enterprises to make more accurate production plans. Therefore, it is of great practical significance to establish a carbon price prediction model with higher prediction accuracy.Based on the existing research results in the field of carbon price prediction in China, this paper takes the carbon price prediction of Beijing, Hubei and Shenzhen carbon exchanges as examples. The main research contents are as follows:

(1) This paper proposes a prediction framework of synchronous multi-decomposition method, and verifies the feasibility of synchronous multi-decomposition method by comparing with the prediction results of a single decomposition method. Firstly, the historical carbon price sequence was decomposed by EMD, OVMD and SSA simultaneously, and the LZ complexity value of each decomposition subsequence was calculated, and cluster reconstruction was carried out according to the elbow rule. The reconstructed sequence was predicted by Ada-Elman model, and nonlinear integration was performed by BP.

(2) This paper proposes a carbon price prediction model based on multiple data information. Firstly, relevant influencing factors were screened according to the existing research literature on influencing factors of China's carbon price, and the data set with the same decomposition subsequence was obtained by input of multiple empirical mode decomposition (MEMD) together with the historical carbon price. The multi-scale walking entropy (MDE) of each subsequence was calculated and reconstructed into high frequency and low frequency. The important influencing factors were selected by Lasso for different frequencies into the prediction model, and the superiority of the model was verified by comparison.

(3) This paper proposes a prediction model that combines multiple decomposition and multiple decomposition methods with Baidu search information. As the largest search engine in China, Baidu's search index can reflect the attention of carbon price to a considerable extent. First of all, the paper selects the Baidu search index with the relevant influential words of carbon price, and makes a GBDT model to rank the importance of the influencing factors of other markets selected by the literature review. Secondly, the selected factors and historical carbon prices are input into multiple empirical mode decomposition and multiple variational mode decomposition (MEMD/MVMD) for decomposition, and the decomposition sequence is reconstructed into high frequency, low frequency and trend terms. Then Pearson correlation coefficient, principal component analysis (PCA) and Granger Causality were introduced to further filter the sequences. Finally, the filtered sequences were input into the convolutional neural network (CNN-BiLSTM) for prediction. The superiority of this model is verified by comparing with other models.

According to the work done in this paper, it can be concluded that: (1) the carbon price prediction based on market influence factors and Baidu search index is reliable and effective; (2) Synchronous multi-decomposition method can improve the prediction effect of carbon price prediction model; (3) When exogenous factor input is added to the carbon price prediction model based on decomposition and reconstruction framework, secondary screening of exogenous factor sequence is required, otherwise the prediction effect of the model will be affected.