基于LSTM NN特征提取的低温气象灾害预报方法研究

Research on Rainfall Forecasting Based on Optimistic Support Vector Machine

针对冬季低温灾害性天气过程具有时间相关性和非线性变化等特点而造成的建模因子处理和预报建模困难问题,以及现有预报方法未能充分获取低温冷害的本质特征,论文提出了一种以深度学习长短期记忆神经网络(LSTM)为基础的低温气象灾害预报模型.该方法结合了长短期记忆神经网络、随机森林和优化支持向量机回归(SVM)等各自的优点,先采用长短期记忆神经网络方法获取数据特征,再利用随机森林方法中的特征重要属性对LSTM模型获取的特征进一步筛选,将随机森林模型获取的特征训练优化支持向量回归建立了低温冷害的预测模型.进一步将该模型与LSTM模型、SVM模型、随机森林回归模型、逐步回归预报模型的预测效果进行了比较,由测试样本的预报仿真实验证明,在相同的预报建模样本和相同的预报因子条件下,该模型比其他4个模型有较高的预测精度,显示了对非线性低温冷害预报问题的更好适用性.

Aiming at the problem that it is difficult to determine the modeling factors and the non-linear weather prediction modeling in weather prediction modeling,a novel method is presented for low-temperature meteorological disaster prediction based on deep learning Long Short Term Memory neural network(LSTM),which the model focuses on how to select factor and how to build a forecast modeling in the winter low temperature disastrous weather process.This method combines the advantages of LSTM,random forests,and optimized support vector machine regression(SVM).Firstly,LSTM neural network is used to obtain data.Secondly,the random forest method is used to extract important modeling information.Lastly,SVR is used to establish the prediction model for low-temperature meteorological disaster prediction modeling.Finally,the prediction effect of the model is further compared with the LSTM model,SVM model,random forest regression model,and stepwise regression forecast model.The computing results show that the present model yields better forecasting performance in this case study,compared to other forecasting models,which model has higher prediction accuracy.Our model may provide a promising alternative for forecasting rainfall application to nonlinear low-temperature cold damage prediction problems.