Research On Weather Elements Prediction Based On Ensemble Learning Model And CNN-ALSTM

The weather of meteorological elements for specific areas in complex terrain is currently a research challenge in the field of meteorology,especially in the case of the Winter Olympics and other outdoor events,where weather conditions are one of the central factors affecting their success.Accurate forecasting of weather elements such as wind speed,temperature,relative humidity and precipitation is directly related to the running of the event,the performance of the competitors and the safety of the events,so the ability to detect and forecast changes in the climate in a timely and accurate manner is crucial to the success of the events.With the advancement of meteorological research,numerical weather prediction models,as the mainstream weather forecasting method,are constantly developing.However,for complex terrain areas,forecast accuracy still needs to be improved,and forecast timeliness is still insufficient to meet actual operational needs.Therefore,it is essential that numerical weather prediction models are followed by revised forecasts.This thesis aims to improve the timeliness and refinement of meteorological forecasts for fixed areas,and to achieve objective revisions of weather elements forecasts based on numerical weather prediction models,combined with actual observations.In this thesis,average wind speed,gustiness speed,relative humidity and precipitation prediction are investigated for a fixed point area with small to medium scale complex terrain,and the main research work and innovation points are as follows.(1)The observation station dataset is designed and constructed.In order to meet the demand for meteorological forecasting in complex terrain areas,this thesis constructs a multifeature dataset for the specified area and completes the extraction of gridded data.The data from the numerical weather prediction model are bilinearly interpolated according to the location information of the observation stations,to convert the data into station data,and historical observation data are extracted under the same time scale to ensure the continuity and consistency of the data.Finally,the features required for the prediction of each meteorological element are filtered out based on meteorological a priori knowledge,combined with machine learning methods.(2)An ensemble learning-based prediction model for weather elements is proposed.The ensemble learning method is applied to develop an objective forecasting technology method for average wind speed,gustiness speed,relative humidity and precipitation for a specified area.The method integrates four individual machine learning models,namely random forest,gradient boosting tree,multiple linear regression and artificial neural network,using a ridge regression model as the ensemble learner.The ensemble learning model effectively integrates the advantages of the four individual machine learners and achieves good revised forecasting results in the forecasting of average wind speed,gustiness speed,relative humidity and precipitation.Comparing predicted and observed data,the wind speed prediction model predicts data within 2m/s of the true value,the root mean square error of the relative humidity prediction model is reduced by about 2Rh,and the precipitation prediction model reduces the root mean square error by about 0.05.At the same time,the ensemble learning model effectively fits the physical significance of each group of features,indicating that the ensemble learning model has a certain degree of interpretability.(3)A weather elements prediction method based on CNN-LSTM with attention mechanism is proposed.Based on the ensemble learning forecasting model,in order to further improve the accuracy of long-time forecasting and to emphasize the correlation between time series,a weather forecasting method that adds an attention mechanism to a combination of convolutional neural networks and long and short-term memory networks is proposed.Finally,the necessity of each improved module is verified through ablation experiments,and the effectiveness of the proposed method is demonstrated by the fact that it still has a high accuracy rate at a long time of prediction,controlling the root mean square error of wind speed within0.031.Analysis of the comparison with the integrated learning model shows that the improved model effectively improves the shortcomings of the integrated learning model in terms of reduced accuracy of long-time forecasts.