【作者】 宋杰；

【导师】 于霞；

【作者基本信息】 沈阳工业大学 ， 计算机科学与技术， 2023， 硕士

【摘要】 随着信息技术的革新和发展,用于气象资料收集方面的投入逐渐增加,气象数据逐渐呈现大数据特征。传统的气象预测方法主要基于统计学,通过建立数学模型描述天气变化,对专业知识和天气初始条件存在依赖,具有一定局限性。基于机器学习的气象预测方法,能高效处理分析海量数据信息,挖掘特征更深层的联系,较为准确地描述天气演变过程,基于此本文建立基于机器学习的降水预测模型,分别对降水量预测和日降水等级分类展开研究。在降水量预测研究中,本文以辽宁省气象局提供的地面观测降水资料为实验对象,构建具有多元时间特征的降水数据。首先使用变分模态分解方法(Variational Mode Decomposition,VMD)降低原始数据的非平稳性。然后采用双向长短期记忆神经网络(Bidirectional Long Short-Term Memory,BiLSTM)处理时序特征。最后采用贝叶斯优化方法(Bayesian Optimization,BO)优化模型参数,建立基于VMD-BO-BiLSTM的降水量预测模型,分别对日降水量、周降水量、15天降水量和月降水量进行训练和测试,与其他模型进行对比,并选取其他站点数据进一步做对比分析。实验结果表明,本文VMD-BO-BiLSTM模型在降水量预测上具有一定优势,鲁棒性更强。并发现降水在长时间尺度上周期性明显。在日降水等级分类研究中,本文以国家环境信息中心提供的辽宁省多站点地面观测降水资料为实验对象,通过输入多个气象要素,基于有监督学习方式进行日降水等级分类研究。首先建立XGBoost模型,然后采用遗传算法(Genetic Algorithm,GA)优化XGBoost的超参数,建立基于GA-XGBoost的日降水等级分类模型,最后与其他基模型对比,并选取不同站点的降水数据进一步对比分析。实验结果表明,本文模型的准确率、召回率、F1值均有一定程度提升,验证了GA-XGBoost模型在日降水等级分类中具有一定优越性和泛化能力。经过上述相关实验分析,对本文基于机器学习的降水预测研究给出展望。更多还原

【Abstract】 With the innovation and development of information technology,the investment for weather data collection is gradually increasing,and weather data is gradually showing big data characteristics.Traditional weather forecasting methods are mainly based on statistics and describe weather changes by building mathematical models,which are dependent on professional knowledge and initial weather conditions and have certain limitations.Based on this,this thesis establishes a precipitation prediction model based on machine learning,which can efficiently process and analyze massive data information,explore the deeper connection of features,and more accurately describe the weather evolution process.In the precipitation prediction study,this thesis uses the ground observation precipitation data provided by the Liaoning Provincial Meteorological Bureau as the experimental object to construct precipitation data with multivariate temporal features.Firstly,the Variational Mode Decomposition(VMD)method is used to reduce the non-stationarity of the original data.Then,Bidirectional Long Short-Term Memory(BiLSTM)neural network is used to process the temporal features.Finally,the Bayesian Optimization algorithm(BO)is used to optimize the model parameters,thus establish a precipitation prediction model based on VMD-BO-BiLSTM,train and test daily,weekly,15-day and monthly precipitation respectively,compare with other models,and select other site data for further comparative analysis.The experimental results show that the VMD-BO-BiLSTM model in this thesis has some advantages on precipitation prediction and has more robustness.And it is found that the periodicity of precipitation becomes more obvious with increasing time scale.In the daily precipitation class classification study,this thesis based on the precipitation data provided by the National Center for Environmental Information for ground observation at several stations in Liaoning Province,daily precipitation class classification study based on supervised learning approach by inputting multiple meteorological elements.Firstly,establish the XGBoost model,then use Genetic Algorithm(GA)to optimize the hyperparameters of XGBoost,establish the daily precipitation class classification model based on GA-XGBoost,and finally compare with other base models and select precipitation data from different stations for further comparison and analysis.The experimental results show that the accuracy,recall,and F1 value of the model in this thesis have been improved to some extent,which verifies the superiority and generalization ability of GA-XGBoost model in daily precipitation class classification.After the above related experimental analysis,an outlook on the research of machine learning based precipitation prediction is given in this thesis.更多还原