基于张量特征-GRU和多头自注意力机制的水质预测模型方法

Prediction Model Based on Tensor Feature-GRU and Multi-head Self-attention Mechanism

针对现有水质预测模型对水质监测指标间存在数据特征利用不足,导致预测精度不高的问题,本文提出了一种基于张量分解融合门控神经网络(gate recurrent unit,GRU)和多头自注意力机制(Multi Head Self-Attention)的多指标水质预测模型(TGMHA)。该模型通过标准延迟嵌入变换(Standard delay embedding transform,SDT)将时序水质指标数据转换为张量数据,利用Tucker张量分解提取数据特征,然后结合多头自注意力机制挖掘多种水质指标数据特征之间的潜在关系,最后采用GRU模型实现多指标水质预测。对比实验证明了该模型预测相比传统GRU水质预测模型,得到的均方根误差(RMSE)、平均绝对误差(MAE)、平均绝对百分比误差(MAPE)和决定系数(R 2)四个指标有5-10%的提升,有效提升了水质预测的精度,具有较好的鲁棒性,为水质预测和环境监测管理提供了科学决策依据。

In order to resolve the problem that existing water quality prediction models were unable to effectively utilize data characteristics between water quality monitoring indicators,which resulted in poor prediction accuracy.Therefore,a multi-index water quality prediction model based on tensor decomposition fusion gated unit(GRU)and self-attention mechanism(Self-Attention)is presented,known as TGMHA.The model converts water quality time series data into tensor data by standard delay embedding transformation(SDT),extracts data characteristics by Tucker tensor decomposition,and then combines multi-head self-attention mechanism to discover potential relationships among data characteristics of multiple water quality indicators.Finally,the GRU model is used to achieve multi-index water quality prediction.Compared with the traditional GRU water quality prediction model,the four indexes of root mean square error(RMSE),mean absolute error(MAE),mean absolute percentage error(MAPE)and determination coefficient(R2)obtained by this model are improved by 5-10%.This method effectively improves the accuracy of water quality prediction and exhibits fine robustness which provides a scientific decision basis for water quality prediction and environmental monitoring management.