|
| 基于ReVIN-TCN-NBEATS混合模型的日碳排放量预测研究 |
| Research on Daily Carbon Emission Prediction Based on ReVIN-TCN-NBEATS Hybrid Model |
| 投稿时间：2026-03-04 修订日期：2026-04-02 |
| DOI： |
| 中文关键词: Research on Daily Carbon Emission Prediction Based on ReVIN-TCN-NBEATS Hybrid Model |
| 英文关键词: carbon emission forecasting daily scale ReVIN hybrid deep learning prediction |
| 基金项目:国家自然科学基金面上项目(52574221)、煤炭安全精准开采国家地方联合工程研究中心（安徽理工大学）开放基金资助(EC2024010)、安徽理工大学高层次引进人才科研启动基金（2024yjrc103）. |
|
| 摘要点击次数: 56 |
| 全文下载次数: 0 |
| 中文摘要: |
| 针对日碳排放数据存在的显著非平稳性、强随机波动及多重周期叠加等典型特征问题。提出了一种融合反向实例归一化（ReVIN）、时间卷积网络（TCN）与神经基扩张分析（N-BEATS）的混合深度学习预测模型（ReVIN-TCN-NBEATS）。该模型首先引入ReVIN机制对输入序列进行双层归一化处理，有效消除了由外部冲击引起的数据分布偏移；随后构建并行特征提取架构，利用TCN捕捉序列局部的非线性高频波动，同时借助N-BEATS拟合全局的长期趋势与周期性规律；最后通过SE-Attention注意力模块根据上下文信息动态分配权重，实现微观波动与宏观趋势的自适应融合。仿真结果表明，基于安徽省2019–2025年日碳排放数据的实证分析表明，该模型在全样本周期内表现出卓越的拟合性能，决定系数（R2）高达0.9737，平均绝对百分比误差（MAPE）仅为0.60%，显著优于LSTM、GRU及单一组件模型。此外，在1至30天的多步预测任务中，该模型始终保持高精度与强稳定性（30天R2仍优于0.96），为实现区域碳排放的精细化监测与“双碳”政策的动态评估提供了精准的技术支撑。 |
| 英文摘要: |
| o address the challenges inherent in daily carbon emission data, such as significant non-stationarity, strong stochastic fluctuations, and multiple superimposed periodicities. This study proposes a hybrid deep learning forecasting model, ReVIN-TCN-NBEATS, which integrates Reverse Instance Normalization (ReVIN), Temporal Convolutional Networks (TCN), and Neural Basis Expansion Analysis (N-BEATS). The model first incorporates the ReVIN mechanism to apply dual-layer normalization to the input sequence, effectively mitigating data distribution shifts induced by external shocks. Subsequently, a parallel feature extraction architecture is constructed: TCN is employed to capture local nonlinear high-frequency fluctuations, while N-BEATS is utilized to model global long-term trends and periodic patterns. Furthermore, an SE-Attention module is introduced to dynamically allocate weights based on contextual information, facilitating the adaptive fusion of micro-level fluctuations and macro-level trends. Empirical analysis using daily carbon emission data from Anhui Province (2019–2025) demonstrates that the proposed model exhibits exceptional fitting performance throughout the entire sample period. It achieves a coefficient of determination (R2) of?0.9737 and a Mean Absolute Percentage Error (MAPE) of only?0.60%, significantly outperforming baselines such as LSTM, GRU, and single-component models. Moreover, in multi-step forecasting tasks ranging from 1 to 30 days, model consistently maintains high accuracy and robustness—retaining an?R2?exceeding?0.96?even at the 30-day horizon. These results suggest that the proposed framework provides precise technical support for the refined monitoring of regional carbon emissions and the dynamic assessment of "dual carbon" policies. |
|
View Fulltext
查看/发表评论 下载PDF阅读器 |
| 关闭 |
分享按钮