Research On The Discovery Method Of Marine Heat Wave Event Pattern Based On Time Series Data

Global marine heatwaves are increasing in frequency and intensity with climate warming.Marine heatwaves are extremely hazardous natural disasters with complex mechanisms that result from the combined effects of internal modes of the climate system and external human activities.The mechanisms of marine heatwaves differ significantly in different marine regions,causing enormous impacts on human society and marine ecosystems.Accurately predicting marine heatwave events is crucial for reducing the damage caused by them.Marine heatwaves are typically defined as extreme events where the sea surface temperature exceeds a fixed threshold,seasonal variation threshold,or cumulative threshold.By mining potential trend patterns hidden in the sea surface temperature time-series data,as well as the superimposed impact patterns of other ocean environmental factors such as wind speed and air pressure,important reference information can be provided for accurate and perceptible heatwave prediction,supporting marine heatwave warning and forecasting in coastal areas.Association rule mining,as an unsupervised machine learning method,is suitable for extracting potential patterns of variation and knowledge models from large amounts of noisy and incomplete marine time-series data,providing support for the modeling of spatial-temporal information and hypothesis generation in marine science.However,traditional Apriori-based association rule mining algorithms face the challenge of a surge in the number of sea surface temperature candidate sequences,leading to an increase in computational complexity.Furthermore,this method requires the discretization of sequence data,which may result in information loss.Therefore,how to propose an effective association rule mining method for marine heatwave event trend patterns based on large-scale sea surface temperature time-series data is an urgent scientific problem that needs to be addressed.In the process of modeling the impact of multiple environmental factors on marine heatwaves,deep learning models exhibit strong learning ability and good generalization ability,making them suitable for modeling the coupled effects of multiple environmental factors when marine heatwaves occur.However,due to the highly nonlinear operations of deep neural network models,their interpretability is poor,and the multi-factor coupling patterns contained in the models cannot be excavated and displayed.The key issue in understanding the causal mechanism of marine heatwaves is how to discover the coupling patterns of multiple environmental factors on marine heatwave events and reveal the degree of influence of other ocean environmental factors on sea surface temperature changes from the deep learning-based marine heatwave prediction model.To address the above-mentioned issues,this thesis proposes a method for discovering patterns of marine heatwaves based on time series data.The main work is as follows:(1)To address the shortcomings of traditional association rule mining techniques applied to the discovery of patterns in sea surface temperature data using a single feature,we propose a method for discovering patterns of marine heatwaves based on constrained motif association rule mining.We introduce motifs to express rules and avoid the problem of information loss caused by the discretization process.Furthermore,we combine the definition of heatwave events with the constraint of the rule mining process to ensure the validity of the mining results.The specific process is as follows: first,the STAMP algorithm is used to discover the most similar subsequence motif in the sequence;then,the motif is divided into the antecedent and consequent of the rule,and the rules are scored using minimum description length.The division method with the highest score is selected,and candidate association rules are obtained based on this method.Finally,the candidate association rules are filtered according to the constraints of the heatwave events to extract trends in marine heatwave patterns.We applied this method to sea surface temperature data from three stations in the nearshore waters of China to conduct experiments on the discovery of patterns in marine heatwaves.The research results show that marine heatwaves in the nearshore waters of China have seasonal characteristics,mainly occurring in summer and lasting for more than 20 days.The patterns of marine heatwave events exhibit regular trends of heating and cooling,with symmetrical heating and cooling rates.Before the occurrence of heatwaves,there is a brief period of warming of the sea surface temperature.(2)In response to the lack of interpretability of deep learning methods in pattern mining tasks,this paper proposes a pattern discovery method based on interpretable deep learning.By quantifying the contribution of ocean environmental factors to the occurrence of heatwave events,the superimposed impact pattern of multiple environmental factors on marine heatwaves is revealed.The specific process is as follows:first,a deep learning model is used to establish a nonlinear predictive relationship between ocean environmental factors(wind speed and air pressure)and sea surface temperature.Then,deep learning interpretable techniques such as Expected Gradients are applied to the trained model to quantify the time-feature importance of ocean environmental factors in predicting heatwave events.Using the Expected Gradients method,feature importance scores for wind speed and air pressure can be obtained and used to analyze the heatwave events identified by relative thresholds,and finally,the superimposed pattern of wind speed and air pressure on heatwave events is explained based on feature importance scores.This paper analyzes 18 heatwave events and reveals two associated patterns between air pressure,wind speed,and heatwave events,corresponding to two mechanisms that trigger marine heatwaves: atmospheric forcing and sea surface wind convergence.