基于LSTM-RNN的海水表面温度模型研究
|
摘要
针对数值模式和统计学习方法在海表面温度(SST)建模中的不足,将长短时记忆循环神经网络(LSTM-RNN)应用于SST的建模。使用研究海区24 a月平均的SST和太阳辐射、风场、蒸发降水等物理参数,通过LSTM-RNN构建西太平洋研究海区SST时间序列变化模型,用于预报研究海区下个月SST。建立了两个模型model1和model2,model1仅使用SST数据作为model2的对照,model2使用SST和其他物理参数。结果表明:model2在验证数据中的MAE为0. 15℃,RMSE为0. 19℃,相关性系数为0. 978,和model1相比总体准确性提升31%,表明LSTM-RNN应用于SST建模是可行的; LSTM-RNN可以建立其他物理参数与SST的关系,从而显著提升海水表面温度模型的准确性。
Due to the shortcomings of numerical modeling and statistical learning methods in SST modeling,this study applies LSTM-RNN( long short term memory recurrent neural network) to improve the SST modeling. Using SST,solar radiation,wind field,evaporation,precipitation and other physical parameters of monthly averaged data of 24 years,the SST time series model of the Western Pacific is constructed by LSTM-RNN to predict the coming month's SST in the study area. Two models,model1 and model2,are established. Model1 only uses SST data as a comparison of model2 that consists of SST and physical parameters. The results show that the MAE of model2 in the valid set is 0. 15℃,RMSE is 0. 19℃ and the correlation coefficient is 0. 978. Compared with model1,the overall accuracy of model2 is higher than 31%. It shows that the application of LSTM-RNN to SST modeling is feasible and LSTM-RNN can get the relationship between physical parameters and SST. Thus,the accuracy of the surface temperature model of sea water can be improved significantly.
Due to the shortcomings of numerical modeling and statistical learning methods in SST modeling,this study applies LSTM-RNN( long short term memory recurrent neural network) to improve the SST modeling. Using SST,solar radiation,wind field,evaporation,precipitation and other physical parameters of monthly averaged data of 24 years,the SST time series model of the Western Pacific is constructed by LSTM-RNN to predict the coming month's SST in the study area. Two models,model1 and model2,are established. Model1 only uses SST data as a comparison of model2 that consists of SST and physical parameters. The results show that the MAE of model2 in the valid set is 0. 15℃,RMSE is 0. 19℃ and the correlation coefficient is 0. 978. Compared with model1,the overall accuracy of model2 is higher than 31%. It shows that the application of LSTM-RNN to SST modeling is feasible and LSTM-RNN can get the relationship between physical parameters and SST. Thus,the accuracy of the surface temperature model of sea water can be improved significantly.
引文
[1]王强.渤-黄-东海海温短期数值预报研究[D].青岛:中国海洋大学,2003:32-43.
[2]张绪东,张菀伦,李云波.北黄海海温分布变化的数值模拟分析[J].海洋预报,2015,32(5):89-97.
[3]笪良龙,熊张浩,过武宏.海洋温度场稳定性与可预报性研究[J].海洋技术学报,2015,34(1):55-61.
[4]张玉荣.大连附近海域幼鲍安全越冬海水温度预报方法的研究[J].海洋环境科学,1992,11(1):34-38.
[5] WU A,HSIEH W W,Tang B. Neural network forecasts of the tropical Pacific sea surface temperatures[J]. Neural Networks,2006,19(2):145-154.
[6]吉进喜,张立凤,郭渊. Kalman滤波技术在海表温度预测中的应用[J].海洋预报,2010,27(3):59-65.
[7] HOCHREITER S,SCHMIDHUBER J. Long short-term memory[J]. Neural Computation,1997,9(8):1 735-1 780.
[8] JASON B. Multivariate time series forecasting with LSTMs in Keras[EB/OL].[2017-08-17]. https://machinelearningmastery.com/multivariate-time-series-forecasting-lstms-keras.
[9]范竣翔,李琦,朱亚杰,等.基于RNN的空气污染时空预报模型研究[J].测绘科学,2017,42(7):76-83.
[10] CHRISTOPHER O. Understanding LSTM networks[EB/OL].[2015-08-27]. http://colah. github. io/posts/2015-08-Understanding-LSTMs.
[11] WAN Z Y,VLACHAS P R,KOUMOUTSAKOS P,et al. Data-assisted reduced-order modeling of extreme events in complex dynamical systems[J]. PLOS ONE,2018,13(5):e0197704.
[12]焦斌,叶明星. BP神经网络隐层单元数确定方法[J].上海电机学院学报,2013,16(3):113-116.
[13] JAIN P,KAR P. Non-convex optimization for machine learning[J]. Foundations and Trendsin Machine Learning,2017,10(3/4):142-336.
[2]张绪东,张菀伦,李云波.北黄海海温分布变化的数值模拟分析[J].海洋预报,2015,32(5):89-97.
[3]笪良龙,熊张浩,过武宏.海洋温度场稳定性与可预报性研究[J].海洋技术学报,2015,34(1):55-61.
[4]张玉荣.大连附近海域幼鲍安全越冬海水温度预报方法的研究[J].海洋环境科学,1992,11(1):34-38.
[5] WU A,HSIEH W W,Tang B. Neural network forecasts of the tropical Pacific sea surface temperatures[J]. Neural Networks,2006,19(2):145-154.
[6]吉进喜,张立凤,郭渊. Kalman滤波技术在海表温度预测中的应用[J].海洋预报,2010,27(3):59-65.
[7] HOCHREITER S,SCHMIDHUBER J. Long short-term memory[J]. Neural Computation,1997,9(8):1 735-1 780.
[8] JASON B. Multivariate time series forecasting with LSTMs in Keras[EB/OL].[2017-08-17]. https://machinelearningmastery.com/multivariate-time-series-forecasting-lstms-keras.
[9]范竣翔,李琦,朱亚杰,等.基于RNN的空气污染时空预报模型研究[J].测绘科学,2017,42(7):76-83.
[10] CHRISTOPHER O. Understanding LSTM networks[EB/OL].[2015-08-27]. http://colah. github. io/posts/2015-08-Understanding-LSTMs.
[11] WAN Z Y,VLACHAS P R,KOUMOUTSAKOS P,et al. Data-assisted reduced-order modeling of extreme events in complex dynamical systems[J]. PLOS ONE,2018,13(5):e0197704.
[12]焦斌,叶明星. BP神经网络隐层单元数确定方法[J].上海电机学院学报,2013,16(3):113-116.
[13] JAIN P,KAR P. Non-convex optimization for machine learning[J]. Foundations and Trendsin Machine Learning,2017,10(3/4):142-336.