结合时间序列分解和神经网络的河流溶解氧预测

为克服小流域数据资料少,河流溶解氧的非平稳特性及动态变化造成的预测困难,提出结合具有自适应噪声的完整集成经验模态分解(CEEMDAN)和Elman动态神经网络的预测方法.使用CEEMDAN方法对原始溶解氧时序数据进行平稳化处理及降噪,提取溶解氧随时间变化的波动特征、周期特征,以及长期趋势,通过计算样本熵(SE)值,将相似的特征序列合并,以减小误差累积,对合并后的新序列分别采用布谷鸟搜索(CS)算法优化的Elman模型进行预测,将各预测值叠加,得到最终预测结果.实验结果表明:CEEMDAN-SE-CS-Elman方法平均绝对误差(E_MA)为0.14;平均绝对百分误差(E_MPA)为2.07%;均方根误差(E_RMS)为0.24;可决系数(R²)达到0.951 6,精度较其他时间序列预测模型有所提高.

Prediction of River Dissolved Oxygen Combined Times Series Decomposition and Neural Network

In order to overcome the difficulties of prediction caused by the unstable characteristics and dynamic changes of river dissolved oxygen in small watershed with few data, a prediction method was proposed based on complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)and Elman dynamic neural network. The CEEMDAN method was used to stabilize and reduce the noise of the original dissolved oxygen time series data. The fluctuation characteristics, periodic characteristics and long-term trend of dissolved oxygen with time change were extracted. By calculating the sample entropy(SE)value, several sequences with similar feature were combined to reduce the error accumulation. The Elman model optimized bythe cuckoo search(CS)algorithm was adopted to predict the new recombined sequences respectively, and thefinal prediction result was obtained by superposing the prediction results of each sequence. The experimental results showed that the mean absolute error(E_MA)of CEEMDAN-SE-CS-Elman method was 0.14, the mean absolute percentage error(E_MPA)was 2.07%, the root mean square error(E_RMS)was 0.24, and the coefficient of determination(R²)reached 0.951 6. The prediction accuracy was improved compared with other time series forecasting models.