Prediction Of Harmful Algal Blooms Based On Machine Learning Technology

As a major marine disaster,harmful algal blooms(HABs)can cause serious damage to marine ecosystems and even endanger human health.As a developing country with both land and sea,China has a coastline of more than 18000 kilometers,especially rich in marine resources,which has important strategic and economic significance for China's development.In recent years,the frequency and scale of harmful algal blooms in China's coastal areas are increasing,which has caused great harm to the marine environment,social economy and even people's health.Therefore,the study of harmful algal blooms is imminent.In view of the high nonlinearity of the outbreak process of marine harmful algal blooms,it is difficult for traditional ecological dynamic models and statistical methods to make accurate prediction.Based on machine learning technology,this paper studies the cause of formation and concentration prediction of harmful algal blooms through characteristics and model selection.The main research contents are described as follows.First,the data from Yantai monitoring center of the North Sea Branch of the National Oceanic Administration and Scripps monitoring station of the United States are selected as the research dataset.Based on the feature selection technology,the full search method is used to analyze all feature subsets and different models.Then based on the GBDT model,the most closely related environmental factors are calculated and analyzed.Secondly,in view of the lack of harmful algal bloom concentration in the experiment,based on the selected feature subset and GBDT model,the missing data of the complete experimental data is filled.Finally,for the completed data,we use GBDT combined with specific input characteristics to predict the phytoplankton concentration one week and two weeks in advance.The simulation results show that this method can basically predict the change trend of phytoplankton concentration one week in advance.In addition,due to different environments,the models with the best performance may be different.In order to further ensure the model generalization,we use the idea of integrated learning for reference,and combine multiple models into meta models based on the Stacking strategy,so that the combined model has strong generalization.In this paper,the method of complete search can ensure to get the best feature subset of the prediction effect,so as to further analyze the relationship between it and the concentration of harmful algal blooms,and provide guidance for the experiment and prevention.Through the selection of characteristics and models,the concentration of harmful algal blooms can be predicted one to two weeks in advance,and then the outbreak of harmful algal blooms can be predicted in advance.The fusion model based on Stacking integration algorithm ensures that the model has strong generalization and learning ability under the experimental data of different locations.