Research On Temperature Forecasting Based On Random Forest And Long Short-term Memory

With the fast burgeon and maturity of computer technology,machine learning has opened up a broad world of its application,involving image recognition,natural language processing and other aspects.At the same time,it also opens up new thinking for the processing and forecasting of massive meteorological data.Since ancient times,meteorological problems have affected all aspects of human life.People have been committed to forecasting of meteorological problems accurately.And,it has more practical research value to accurately forecast the temperature over the next several days and accurately capture the dynamic changes of future meteorology.Meteorological system is a complex system affected by many factors.Most meteorological series have same features like time series.Through the sorting,cleaning,filling and analysis of the meteorological,we find that there are seasonal and periodic laws in the data.Meanwhile,there is auto-correlation in the time series data.Compared with the traditional single factor time series forecasting models,such as SARIMA,this paper takes two multi factors forecasting models as the research methods: long short-term memory neural network and random forest.Then we calculate the correlation coefficient between each influencing factor and temperature,and select several factors with strong correlation with temperature combined with scatter diagram.Aiming at the problem that the seasonal feature of the series is not considered,it is considered to quantify the seasonal feature in the series into numerical variable by One-Hot encoding.The experiments show that the two models with seasonal variable have better performance in forecasting.At the same time,because many super parameters are involved in the construction of random forest,two search algorithms are proposed to optimize the parameters to ensure the optimization of the model.Furthermore,aiming at the autocorrelation of temperature series,LSTM model is constructed,and a sliding window is set to ensure that the network can learn enough information.In addition,in order to avoid over fitting and gradient explosion,the early-stopping mechanism is introduced into the LSTM layer.The results show that the improved LSTM always has high forecasting accuracy and quick convergence speed.To compare the forecasting ability of several improved models,short-term forecasting and medium-term forecasting experiments are carried out respectively.The experiments results show that the error in short-term forecasting is smaller for each model.And the result is closer to the actual value.It means that the improved LSTM has better fitting effect than other models.And the forecasting error is about 3% lower than other models.