基于LSTM和FNN的昆明市气候舒适度相关气象指标预测方法

Prediction method of related meteorological indexes of Kunming climate comfort based on LSTM and FNN

针对昆明市缺乏较好区域气候舒适度预测模型、气象指标预测方法性能不佳的问题,结合长短期记忆(LSTM)网络的时间序列预测方面和前馈神经网络(FNN)的数据整合以及增强相关性的优势,提出一种基于LSTMFNN模型的昆明市气候舒适度相关气象指标预测方法;并根据适宜旅游区域随季节变化而变化的特点,提出季节划分下的气候舒适度相关气象指标预测方法。利用昆明市气象站1981—2010年共30年的平均数据,构造273 d气象数据的训练集和92 d气象数据的测试集对LSTM-FNN预测模型进行实验,日平均气温、相对空气湿度和日均风速三项指标的预测数据与真实数据对比所得平均绝对误差(MAE)分别是0.14℃、0.45%、0.13 m/s;LSTM-FNN预测模型与实验参数下的原始LSTM、双向长短期记忆(BiLSTM)网络和双向循环神经网络(BRNN)相比,日均气温MAE分别降低了0.05℃、0.27℃、0.09℃,其余两项气象指标预测性能基本一致。

In view of the lack of a good regional climate comfort prediction model,the poor performance of the weather index prediction method for Kunming,combined with the advantages of the time series prediction of Long Short-Term Memory(LSTM)network and the data integration and enhanced correlation of of Feedforward Neural Network(FNN),a method for predicting climate comfort related meteorological indexes in Kunming based on the LSTM-FNN model was proposed;and according to the characteristics of seasonal changes in suitable tourist areas,the climate comfort related meteorological indexes prediction method under seasonal division was proposed. By using the 30-year average data of Kunming Meteorological Station from 1981 to 2010,a training set of 273 d meteorological data and a test set of 92 d meteorological data were built for experiment of the LSTM-FNN prediction model,including daily average temperature,relative air humidity and daily average wind speed. The Mean Absolute Error(MAE)obtained is 0. 14℃,0. 45%,and0. 13 m/s. Compared with the original LSTM,Bi-directional LSTM(BiLSTM)network and Bi-directional Recurrent Neural Network(BRNN) under the experimental parameters,the LSTM-FNN prediction model reduces the daily average temperature MAE by 0. 05℃,0. 27℃,and 0. 09℃,respectively,and the prediction performance of the other two meteorological indexes is basically the same.