基于双阶段注意力权重优化机制的LSTM地表水溶解氧预测模型研究
A Study on Dissolved Oxygen Prediction Based a LSTM Model with Improved Weights Dual-stage Attention Mechanism
-
摘要: 溶解氧是水体中的重要水质指标,准确预报其变化规律对水环境管理具有重要意义.在提出一种基于双阶段注意力权重优化机制的LSTM地表水溶解氧预测模型(DAIW-LSTM Model)的基础上,将新模型应用于流溪河流域水质监测站溶解氧日均值预测研究.该模型在编码器使用了空间注意力,解码器使用了时间注意力,编码器和解码器各自包含了两个阶段的权重优化的新机制.鉴于新模型在对非线性更强、非平稳性更突出的时间序列数据预测的优势,开展了三个水质监测站不同基线模型预测效果的对比分析,探讨了上游站点特征输入及特征权重优化机制对模型预测性能的影响.研究结果表明:①通过与DA-LSTM、LSTM、Bi-LSTM等基线模型的对比,验证了模型适用性和精准性;DAIW-LSTM模型对于白云李溪坝站溶解氧预测的SMAPE、MAE、MSE三个指标分别为0.075、0.611、0.712是所有模型中最优的,相比DA-LSTM模型分别降低了9.61%、8.06%、18.17%.②进一步加入上游特征的输入,新模型的预测效果进一步提升,该实验也证明上游特征变量选取的重要性;另外,注意力机制引入,有效解决了传统LSTM模型在额外噪音输入情形下鲁棒性不足的问题.③对于新的注意力权重优化机制,二阶段会对一阶段的初步权重进行优化修正,针对pH、电导率、水温、气温等影响溶解氧预测的重要特征,新模型会在特定的时候自动强化其影响权重,从而提高模型的预测精度.研究结果对于地表水水质预测研究具有较好的借鉴作用.Abstract: Dissolved Oxygen(DO) is a key index of the water environment. Accurate prediction of time series of DO concentrations is greatly important for water environment management. A novel DO concentrations prediction model based on LSTM method with improved weights dual-stage Attention mechanism (DAIW-LSTM Model) was proposed, and was applied to predict the daily DO concentrations in Liuxi River Basin. The model uses spatial attention in encoder and temporal attention in decoder, the encoder and decoder each contain a new mechanism of weight optimization in two stages. Taking the advantages of the new model in predicting time series data with stronger nonlinearity and more prominent non-stationarity, a comparative analysis among different baseline models for three water quality monitoring stations was carried out, and the upstream feature variables input and the effects of feature weight optimization mechanism were discussed. Results showed that: (1) Comparing with baseline models such as DA-LSTM, LSTM, Bi-LSTM, the applicability and accuracy of the model were verified; The SMAPE, MAE and MSE predicted by DAIW-LSTM model at Baiyunlixiba station were 0.075, 0.611 and 0.712 respectively, which were the best of all models, and could be reduced by 9.61%, 8.06% and 18.17% respectively compared with DA-LSTM model. (2) Further tests with the input of upstream characteristics showed that the prediction effect of the new model was further improved, it also proved the importance of selecting upstream feature variables. The attention mechanism could effectively solve the problem of insufficient robustness of the traditional LSTM model with additional noise. (3) The second stage could optimize and correct the initial weight of the first stage in the new attention weight optimization mechanism. The prediction accuracy of the new model would be improved, since the important features such as pH, conductivity, water temperature, and air temperature, were automatically strengthened. The research results could be used as a good reference for the research on water quality predictions.
-
Key words:
- Attention mechanism /
- Time series prediction /
- Dissolved oxygen prediction /
- LSTM model
点击查看大图
计量
- 文章访问数: 209
- HTML全文浏览量: 20
- PDF下载量: 26
- 被引次数: 0