基于VMDLSTNet的水质预测模型

Water Quality Prediction Model Based on VMDLSTNet

针对水质时序预测中存在长期信息和短期信息混合导致预测精度低的问题,采用变分模态分解(variational mode decomposition, VMD)和长短期时间序列网络(long-and short-term time-series network, LSTNet)组合使用以期望解决该问题得出更准确的水质预测。LSTNet网络中使用卷积神经网络(convolutional neural networks, CNN)提取短期局部水质信息,使用循环神经网络(recurrent neural network, RNN)提取长期水质信息,并且通过Skip-RNN利用序列周期特性,提取更长期信息,同时模拟自回归(autoregressive model, AR),为水质预测增添线性成分来达到输出能够响应输入尺度变化的目的。采用珠江流域老口站隔日采样的溶解氧数据验证模型效果,结果表明,VMDLSTNet网络处理水质预测问题的能力,不仅优于传统的BP神经网络(back propagation neural network, BPNN)、支持回归机(support vector regression, SVR)模型,而且优于深度学习中时域卷积网络(temporal convolutional network, TCN)模型、门循环网络(gate recurrent unit, GRU)、增加注意力机制的长短时记忆网络(long short-term memory add attention, LSTM-AT)模型,溶解氧的预测平均绝对误差(mean absolute error, MAE)为0.093 1,预测均方误差(mean square error, MSE)为0.014 6,预测均方根误差(root mean square error, RMSE)为0.120 8,水质类别的预测准确率为95%。

In order to solve the problem of low prediction accuracy caused by the combination of long-term and short-term information in water quality time series prediction, a more accurate water quality prediction is obtained by combining variational mode decomposition(VMD) and long-and short-term time-series networks(LSTNet). Convolutional Neural Networks were used in the LSTNet to extract short-term local water quality information, while recurrent neural networks(RNN) were used to extract long-term water quality information, and Skip-RNN was used to extract even more long-term information by using sequence cycle characteristics. Autoregressive models were also simulated to add linear components to water quality prediction, allowing the output to respond to changes in input scale. Autoregressive models are also simulated to add linear components to water quality prediction, allowing the output to respond to changes in input scale. In terms of water quality prediction, the results show that VMDLSTNet outperforms the traditional back propagation neural network(BPNN) and support vector regression(SVR). Furthermore, it outperforms the temporal convolutional network(TCN) model, gate recurrent unit(GRU) model, and short and long-term memory network model of deep learning attention-enhancing mechanism. The mean absolute error in predicting dissolved oxygen is 0.093 1, and the mean square error is 0.014 6. The root mean square error of prediction is 0.120 8, and the water quality prediction accuracy is 95%.