卷烟生产企业的碳排放量预测研究

Carbon Emission Forecasting Research for Cigarette Manufacturing Enterprises

摘要: 在“双碳”战略持续推进下，能源密集型制造业对碳排放预测的准确性和时效性提出了更高要求。卷烟生产过程中碳排放受多因素耦合与时间波动影响，传统线性模型和部分机器学习方法难以充分捕捉其非线性特征和长短期依赖关系。为此，研究构建了一种融合双向长短期记忆网络和注意力机制，并由蜣螂优化算法(DBO)调优的预测模型(DBO-BiLSTM-Attention)。该模型通过双向序列建模与关键时间步动态加权提升特征表达能力，利用DBO进行全局超参数搜索，增强了模型稳定性与泛化性能。实验基于华东某卷烟企业2022~2024年数据，采用滑动窗口进行多变量时间序列预测。结果表明，该方法相较TCN-LSTM和Attention-LSTM模型，MAE降低约30%，R²提高至92.40%，在预测精度和波动敏感性方面均表现优异，为卷烟生产企业的碳排放动态管理提供了技术支撑。

Abstract: Under the ongoing advancement of the “dual carbon” strategy, energy-intensive manufacturing industries face growing demands for the accuracy and timeliness of carbon emission forecasting. In cigarette production, carbon emissions are influenced by multiple factors and time-varying patterns. Conventional linear models and some machine learning approaches often fail to capture such nonlinear characteristics and temporal dependencies. To address these challenges, a prediction model integrating Bidirectional Long Short-Term Memory networks and an attention mechanism was developed and further optimized using the Dung Beetle Optimization (DBO) algorithm (DBO-BiLSTM-Attention). The model uses bidirectional sequence modeling and dynamic weighting of key time steps to enhance feature representation. Additionally, the Dung Beetle Optimization algorithm performs a global hyperparameter search, improving stability and generalization performance. Experiments were conducted using production data collected from a cigarette manufacturing enterprise in East China between 2022 and 2024, applying a sliding window approach for multivariate time series forecasting. Results indicate that compared with TCN-LSTM and Attention-LSTM models, the proposed method reduced MAE by approximately 30% and increased R² to 92.40%, demonstrating superior predictive accuracy and sensitivity to fluctuations. These findings provide robust technical support for the dynamic management of carbon emissions in cigarette production enterprises.