基于CEEMDAN-Transformer的灌浆流量混合预测模型

灌浆流量是最重要的水利工程灌浆参数之一,通过对灌浆流量的有效预测,可以实现对异常工况的提前响应,以保障施工质量与工程安全。然而由于灌浆过程面临的复杂地质情况,灌浆流量数据存在强非线性与波动性的特点,难以获得令人满意的计算精度。现有灌浆流量预测存在的不足如下:传统神经网络模型对时间序列特征提取和加工处理不足,导致预测精度有限;传统神经网络模型测试集进行一次计算仅能输出一个结果,进行多个时间步预测需要繁杂的多次计算;单测点预测结果预测时间短并且无法反映灌浆流量序列变化的整体趋势,不利于控制灌浆流量和保障施工质量。针对上述问题,本研究提出基于CEEMDAN-Transformer的灌浆流量混合预测模型。基于完全自适应噪声集合经验模态分解(Complete Ensemble Empirical Mode Decomposition with Adaptive Noise,CEEMDAN)方法将灌浆流量分解为本征模函数与残差信号,解决灌浆流量数据的非线性与强波动的问题;采用多头注意力Transformer实现多个本征模函数(Intrinsic Mode Function,IMF)序列到序列的预测,采用多头注意力机制来构建输入和输出的全局依赖关系,提升时间序列参数特征提取水平;最后,建立时序测点多输入多输出模型实现灌浆流量预测,提升多输出序列计算效率,反映整体趋势的多输出序列能够为灌浆流量控制提供参考。工程应用结果表明,本研究提出的基于CEEMDAN-Transformer的灌浆流量混合预测模型具有较好的计算精度和计算效率。

Grouting flow hybrid prediction model based on CEEMDAN-Transformer

Grouting flow is one of the most important grouting parameters of hydraulic engineering.The abnormal construction condition can be found by the effective grouting flow prediction to guarantee the construction quality and safety.However,the geological condition is complex and grouting flow data has the features of strong nonlinearity and volatility,therefore the prediction precision is unsatisfied.The shortcomings of the existing grouting flow prediction are as follows:the traditional neural network model is insufficient in feature extraction,resulting in unsatisfied prediction precision;the traditional neural network model calculates one result by one calculation,multiple time step prediction requires complex multiple calculations;the prediction time of one point is short and the prediction result can not reflect the total trend of grouting flow sequence,therefore it is not beneficial to control grouting flow and guarantee construction quality.For those problems,this research proposes the grouting flow hybrid prediction model based on CEEMDAN-Transformer.The grouting flow is decomposed to eigenmode function and residual signal based on Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN),and the problems of strong nonlinearity and volatility are settled.The sequence prediction of Intrinsic Mode Function (IMF) is realized using multi-head attention Transformer,and the total dependency between input data and output data is established using multi-head attention method.This method is effective in extracting dynamic temporal features and improving the extracting quality.Finally,the grouting flow prediction model with multi-input and multi-output is established to improve the calculation efficiency,providing the reference for grouting flow control.The proposed CEEMDAN-Transformer model has better calculation accuracy and efficiency in grouting flow prediction.