Research And Application Of Multiple Meteorological Data Methods

High-quality meteorological data information is an important basis for conducting research on various meteorological services.For example,the new global concept of smart cities and green energy-efficient buildings based on meteorological data is proposed.The research of meteorological data at home and abroad also involves many fields,such as meteorological data clustering,prediction,and visualization research.This thesis focuses on prediction model of meteorological data.Many scholars have studied meteorological data.However,there are still some key challenges.This thesis summarizes the three key challenges of multivariate meteorological index prediction modeling.In response to the above-mentioned key challenges,this thesis proposes a set of multi-meteorological data research system framework called DA_GRU.The framework divides the meteorological research into three sub-algorithm modules:data preprocessing,group meteorological sites and meteorological index prediction.Specifically,the key challenges and solutions are follows:(1)How to ensure high-quality meteorological data information?This article performs data preprocessing on the original meteorological data.Specifically,the first algorithm modules is data preprocessing which including based on Kronecker Compressed Sensing technology to fill the missing data,and based on the Variational Auto-Encoder to detect the abnormal meteorological data.After the above-mentioned data preprocessing operation,this thesis obtain a complete meteorological dataset without abnormal data,which lays a good data foundation for later meteorological modeling.(2)How to resolve the impact of geographic differences when predict the meteorological data?Considering China has a vast area and has different geographical and climatic conditions,researcher cannot assume that all weather stations have the same weight,when constructing a multivariate meteorological prediction model.Therefore,this thesis uses the dynamic time warping similarity calculation method to divide China's 390 meteorological stations into six group meteorological stations which has the most similar climatic conditions.(3)How to map the weight ratios of different historical sampling time and local seasonality in designing forecasting model?In each group of meteorological station clusters,this thesis fully considers the historical data information of meteorological indicators and the impact of the driving meteorological indicator on the current target indicator to design a dual attention mechanism predict model.Specifically,the model includes an input attention mechanism and a temporal attention mechanism in the encoding-decoding stage,which can adaptively extracting the feature of meteorological data and computes the attention weights based on the previous decoder hidden state that can represent the input information as a weighted sum of the encoder hidden states across all the time steps.In addition,the model also maps the local seasonal relationship of the meteorological data and contains the AR autoregressive linear calibration module.Then,based on Matlab and Tensorflow technologies,this thesis implement the DA_GRU model.In the fifth part,this thesis conduct several independent experiments based proposed DA_GRU model,and has different model performance comparison,analysis and discussion in some traditional meteorological data study algorithms such as AR,LSTM,and RNN.The experimental results show that the DA_GRU research model presented shows better competitiveness,and the predicting errors of RMSE,MAE,and MAPE will be 3.83,1.99 and 0.52 lower,respectively.It is worth emphasizing that this thesis also analyzes and discusses the performance of meteorological indicators under different missing conditions,the effectiveness of anomaly detection models,and the performance comparison of similarity measurement methods of different meteorological stations.The experimental results prove that the missing data processing method proposed in this thesis has lower reconstruction error and strong universality.The anomaly detection model and the design model of the group weather station also have more objective performance,which improves the prediction model,which means that the design of the two sub-modules of data preprocessing and group meteorological station construction in this thesis has played a positive role in the multivariate meteorological prediction model designing.In addition,this thesis also has experiments performed on four open source datasets of traffic flow,solar power generation,power load,and financial exchange rate.The experimental results show that the DA_GRU model designed in this thesis has acceptable universality for others time series data.