摘要
针对火电机组宽负荷动态调节下功率预测精度不高的问题,该文提出一种多模态分解与深度学习协同优化的混合预测模型.模型采用改进型ICEEMDAN对多源传感器信号进行多尺度解耦,结合样本熵引导的KMeans聚类构建频率特征子空间,并利用VMD对高频噪声进行二次降噪.引入冠豪猪优化算法动态调整CNN-BiLSTM网络的正则化系数、初始学习率与隐层节点数,实现预测性能最优化.在山东某机组实测数据上,模型在变工况测试集上将MAE降低至19.45 MW,较传统方法提升20%.结果表明,该双级分解—聚类—优化策略有效缓解了模态混叠与参数敏感问题,为火电机组智能调峰提供了高适应性的预测方法,相关模型已成功集成至某电厂的实时监控系统中,在实际运行中展现出良好的预测准确性与响应速度,显著提升了机组在动态负荷条件下的调峰决策效率与运行经济性.
This paper proposes a hybrid forecasting model based on multi-modal signal decomposition and deep learning collaborative optimization.The model employs an improved complete ensemble empirical mode decomposition with adaptive noise(ICEEMDAN)to achieve multi-scale decoupling of multi-source sensor signals.A frequency feature subspace is then constructed via sample entropy-guided K-means clustering,followed by variational mode decomposition(VMD)to decrease high-frequency noise components secondary.Furthermore,the Capybara Optimization Algorithm(CPO)is introduced to dynamically tune the regularization coefficients,learning rate,and hidden layer structure of the CNN-BiLSTM network,thereby optimizing its predictive performance.Validated on real-world operational data from a thermal unit in Shandong Province,the proposed model reduces the mean absolute error(MAE)on variable condition test sets to 19.45 MW,achieving a 20%improvement over conventional methods.The results demonstrate that the dual-stage strategy of decomposition-clustering-optimization effectively mitigates modal aliasing and parameter sensitivity issues,offering a highly adaptive prediction approach for intelligent peak regulation of thermal units.The model has been successfully integrated into the real-time monitoring system of a power plant,significantly enhancing decision-making efficiency and operational economy under dynamic load conditions.
作者
王杰
闫峻熙
黄刘松
黄莺
张飞
周明琴
WANG Jie;YAN Junxi;HUANG Liusong;HUANG Ying;ZHANG Fei;ZHOU Mingqin(Ma'anshan Teacher's College,Ma'anshan 243041,China;Jilin Provincial Experimental School,Changchun 130000,China;Guodian Nanjing Automation Co.,Ltd,Nanjing 213017,China)
出处
《通化师范学院学报》
2025年第8期1-8,共8页
Journal of Tonghua Normal University
基金
安徽省高校自然科学研究重大项目(2022AH040346)
安徽省人才项目(2024jsqygz150)。
关键词
多模态信号分解
火电功率预测
冠豪猪优化算法
深度双向网络
工况特征解耦
multi-modal signal decomposition
thermal power prediction
CPO
CNN-BiLSTM network
operating condition feature decoupling
作者简介
王杰,安徽固镇人,马鞍山师范高等专科学校教授。
