基于时间卷积神经网络和双尺度特征选择的混合碳价预测模型  

Hybrid Carbon Price Prediction Model Based on Time Convolution Neural Network and Double-scale Feature Selection

作　　者：周建国[1] 韦斯悌 ZHOU Jianguo;WEI Siti(Department of Economics and Management,North China Electric Power University,Baoding 071003,China)

机构地区：[1]华北电力大学经济管理系,河北保定071003

出　　处：《电力科学与工程》2023年第4期41-49,共9页Electric Power Science and Engineering

摘　　要：针对传统碳价格预测模型存在的过拟合和无法有效提取相关特征的问题,提出了一种混合预测模型。首先,通过改进的完全自适应噪声集成经验模态分解算法对原始序列进行分解,以降低数据的波动性和复杂性;然后,用模糊熵对剩余子序列进行重构;此后,利用偏自相关函数和随机森林对子序列进行双尺度特征选择,确定最佳输入维度,以减少不相关特征的输入;最后,通过时间卷积网络进行预测。实验结果表明,与对比模型相比,所提出的模型具有优越性和鲁棒性。该研究结果可为碳市场发展和减排路径相关研究提供有意义的参考。Aiming at the problems of over-fitting and inability to extract relevant features of traditional carbon price prediction models,a hybrid prediction model was proposed.Firstly,the original sequence is decomposed by the improved fully adaptive noise ensemble empirical mode decomposition algorithm to reduce the volatility and complexity of data.Then,the remaining subsequences are reconstructed by fuzzy entropy.After that,partial autocorrelation function and random forest are used to select two-scale features of subsequences,and the best input dimension is determined to reduce the input of irrelevant features.Finally,the prediction is made by time convolution network.The experimental results show that the proposed model is superior and robust compared with the comparison model.The research results can provide meaningful reference for relevant researches of carbon market development and emission reduction path.

关 键 词：碳价格预测 双尺度特征选择 序列重构 时间卷积神经网络 

分 类 号：TK-9[动力工程及工程热物理]