Research On Weather Forecasting Based On Deep Learning

Following the steps of rapidly-developed computer technology,deep learning has opened up a new era of artificial intelligence.On behalf of the new technology innovation,the deeplearning gains breakthrough progress in computer vision,speech recognition,natural language processing and etc.Established in that,the application of deeplearning is not only a challenge,but also bringing an opportunity to develop the meteorological forecasting technology.This paper was modeled from the time series for the temperature sequence,By analyzing the current situation of research at home and abroad and comparing various kinds of forecasting models based on time series,the idea of transforming the depth learning framework to predict the temperature time series was put forward.Considering the weather parameters in the traditional neural network are deemed to be independent with each other and the relationship between them is generally ignored,the means of ?sliding time window? was put forward in the construction of the Deep Feedforward Network weather forecasting model,so that the common neural network can also get the good characteristics of historical time series.It can be found that in the prediction of the temperature,the experimental results taking into account the factor of time series are obviously better than those without consideration of the factor in the time prediction model proposed by the Deep Feedforward Network.Furthermore,in terms of the shortcoming appeared in the Deep Feedforward Network of the experiment that the prediction accuracy declines rapidly with the prolonging of forecasting time,the temperature prediction model for constructing recurrent neural network and the long short-term memeory(LSTM)model specifically used for solving the issue of Long-time Dependency were put forward.After the analysis of technical characteristics of the recurrent neural network and the RNN-LSTM and RNN-GRU network,this study had repeated trial experiments with the consideration of practical problems such as over-fitting,gradient vanishing and gradient exploding problem.Then the strategy that adding various kinds of Regularization optimization and applying Re LU activation function was put forward.This strategy allowed the improved temperature prediction model to be converged in the short time.Finally,operations such as the transformation,cleaning,attribute selection and feature extraction of the meteorological data were also included in the current experiment.As for the platform application,the experiment was moved to Tensor Flow-GPU,Google's latest neural network framework of deep learning.Also with the use of GPU in parallel operation,it is more possible for the more complex experiment with deeper model to be realized.In order to verify the performance of the model,the comparison between the deep learning frameworks in the experiment was carried out and the ARIMA model was added.In this paper,the application research of deep learning technology in the fine prediction of weather was proposed.This attempt provided a brand new way to study the field and explore fine weather forecasting methods.