优化NARX神经网络对时间序列溶解氧的预测

Time-series dissolved oxygen prediction based on optimized NARX neural network

为提高非线性有源自回归(NRAX)神经网络模型的预测精准度,采用主成分分析(PCA)法和灰色关联分析(GRA)法提取原始数据特征并减少输入变量的维度,通过构建PCA-NARX和GRA-NARX模型预测地表水体未来短期(48 h)溶解氧(DO)的质量浓度。结果表明:GRA-NARX模型对时间序列DO质量浓度的预测效果优于NARX模型和PCA-NARX模型,然而预测精度随预测时间的推移呈下降趋势,但是短期预测效果较好,36 h内预测误差可控制在-0.5~0.5 mg/L以内,预测均方根误差和平均绝对误差分别为0.261 mg/L和1.98%。GRA-NARX模型对DO质量浓度预测精度较好,可结合DO与其他水质指标之间的相关性分析,为地表水体水质预测预警和应急响应提供技术支撑。

To improve the prediction accuracy of the nonlinear autoregressive with exogenous input(NARX)neural network,principal component analysis(PCA)and grey relation analysis(GRA)were used to extract features and reduce dimensions of input variables.PCA-NARX and GRA-NARX neural network models were established to predict the change of future short-term(48 h)dissolved oxygen(DO)mass concentration of surface water.The results show that GRA-NARX model has a better prediction performance of time-series DO mass concentration than NARX and PCA-NARX models.However,the prediction accuracy tends to decrease with the prediction time,but the short-term prediction is better,and the prediction error can be controlled within-0.5-0.5 mg/L within 36 h.The root mean square error and mean absolute percentage error of prediction are 0.261 mg/L and 1.98%respectively.The method has good prediction accuracy for DO mass concentration,and can be combined with the correlation analysis between DO and other water quality indicators to provide technical support for surface water quality prediction and early warning and emergency response.