多重注意机制及权重校正LSTM的PVC含水率预测

针对PVC干燥工段中，PVC含水率存在非线性、大滞后、与其他变量关联性复杂难以预测的问题，提出一种多重注意机制及权重校正型长短期记忆网络(LSTM)模型用于PVC含水率的预测。在编码器部分，利用与含水率相关的输入序列之间的相关性对空间注意机制训练的可变权重进行校正，避免由于单纯数据训练导致相关性强的输入变量之间权重差异较大，进而实际干燥工艺不符；同时，由于含水率预测的滞后特性，为减弱长子时间窗口内LSTM单元细胞状态信息丢失，提出信息补偿机制补偿之前时刻细胞状态信息。在解码器部分，利用时间注意机制对编码器的隐藏层状态进行权重更新，解除固定长度向量对模型性能的限制。最后，选取某化工公司干燥工段DCS数据进行验证，相对于RNN、VA-LSTM、STA-LSTM相关系数(R²)分别提高了571%、122.6%、82.6%,结果表明本文模型具有一定优越性。

Prediction of PVC moisture content by multiple attention mechanism and weight correction LSTM

In view of the problems of PVC moisture content in the PVC drying section, such as nonlinearity, large lag, complex correlation with other variables and difficult to predict, a multiple attention mechanism and weight correction long-term and short-term memory network (LSTM) model are proposed for the prediction of PVC moisture content. In the encoder part: use the correlation between the input sequences related to water content to correct the variable weight of spatial attention mechanism training, so as to avoid the large weight difference between the input variables with strong correlation due to simple data training, and then the actual drying process is inconsistent. At the same time, due to the hysteresis of water content prediction, in order to reduce the loss of cell state information of LSTM unit in the eldest son time window, an information compensation mechanism is proposed to compensate the cell state information at the previous time. In the decoder part, we use the time attention mechanism to update the weight of the hidden layer state of the encoder, and remove the limitation of the fixed length vector on the performance of the model. Finally, the DCS data of the drying section of a chemical company were selected for verification. Compared with RNN, VA-LSTM and STA-LSTM, the correlation coefficient (R2) were increased by 571%, 122.6% and 82.6% respectively. The results showed that the model in this paper had certain advantages.