基于改进CEEMDAN的深度学习电煤库存STIFM

Deep Learning Coal Inventory STIFM Based on Improved CEEMDAN

准确预测燃煤电厂的电煤库存是优化能源储存、保障电力供应的重要依据。针对实际生活中短期电煤库存呈不平稳性、随机性和局部突变等特点,提出一种基于改进CEEMDAN分解的TCN-BiGRU-Attention组合模型电煤库存短期预测方法,分析电煤库存特征并选取主要影响因素,将影响因素通过词向量的方式构建成新时序序列,利用于完全自适应噪声集合经验模态分解(Complete EEMD with Adaptive Noise, CEEMDAN)分解数据后通过零率(Zero Crossing Rate, ZCR)将分量分类为高、中和低频并叠加求和,通过时序卷积网络(Temporal Convolutional Network, TCN)提取不同频段时序序列的隐藏特征,以特征向量的方式输入双向门控循环神经网络(Bidirectional Gated Recurrent Unite, BiGRU),并结合Attention机制(Attention Mechanism)给予不同权值突出关键特征并产生预测结果,将各频段序列预测结果求和产生最终预测结果。结果表明,上述模型比单一和其它组合模型预测结果更准确。

Accurate prediction of thermal coal inventory in coal-fired power plants is an important basis for opti⁃mizing energy storage and ensuring power supply.In view of the characteristics of short-term thermal coal inventory in real life,such as non-instability,randomness and local mutation,a short-term prediction method of thermal coal in⁃ventory based on TCN-BiGRU-Attention combination model based on improved CEEMDAN decomposition is pro⁃posed,the characteristics of thermal coal inventory are analyzed and the main influencing factors are selected.The in⁃fluencing factors are constructed into a new time sequence by the way of word vector,which is used in Complete EE⁃MD with Adaptive Noise.After decomposing the data,CEEMDAN classifies the components into high,medium and low frequencies by Zero Crossing Rate(ZCR)and superpositions them and sums them,and extracts the hidden fea⁃tures of the Temporal Convolutional Network(TCN)in different frequency bands.Input Bidirectional Gated Recurrent Unite(BiGRU)in the form of feature vectors,and combine with Attention Mechanism to give different weights to highlight key features and generate prediction results.The final prediction result is obtained by summing the prediction results of each frequency band sequence.The results show that the model is more accurate than the single and other combined models..