Earthquake Prediction Based On Anomaly Detection Algorithm Of Hot Spring Water Chemistry

Earthquake prediction is a world problem that has not yet been solved and numerous studies have shown that earthquake precursors are present in many earthquake events,which are an important area of earthquake research.Among them,seismic fluid geochemical anomalies are one of the most common earthquake precursors.Many studies have shown that there is a high spatial and temporal correlation between seismic fluid geochemical anomalies and earthquake events,which shows great potential in earthquake prediction.However,the method widely used in current research is still based on traditional statistical analysis to manually discriminate anomalies in monitoring data,with problems of high workload,low efficiency and poor accuracy,and the forecasting effect often fails to meet the actual needs.In recent years,with the continuous development of seismic fluid monitoring,a huge amount of monitoring data has been accumulated,which provides a basis for conducting machine learning research and establishing more efficient data processing and earthquake prediction methods.Therefore.Establishing an earthquake prediction model based on machine learning algorithm applicable to seismic fluid geochemistry and carrying out earthquake prediction research are of great theoretical and practical significance for understanding the mechanism of earthquake precursor occurrence,improving earthquake prediction methods,enhancing the efficiency and accuracy of earthquake prediction,and improving seismic fluid geochemistry monitoring schemes.In this study,the active tectonic zone of the southeast coast of China was taken as the study area.Basing on long-term monitoring data of hot spring water chemistry,machine learning algorithms were used to conduct research on the detection of anomalies in hot spring water chemistry,established a new seismic prediction model,tested the prediction performance of various algorithms,and analyzed various factors affecting the predictive performance of the model.The main work and results are as follows.（1）Based on long-term monitoring of hot spring water chemistry,the hydrogeochemical characteristics of six hot springs were obtained,and the reservoir temperature and hot water circulation depth were estimated.The results show that mixing between different water bodies is common those hot springs.Among them,a proportion of seawater is mixed in Gangwei Hot Spring and Shantou Hot Spring,forming Cl-Na·Ca and Cl-Na,respectively;there is shallow groundwater mixing in Tadou Hot Spring,Longmen Hot Spring,Tainai Hot Spring and Gui’an Hot Spring,and hydrochemical type of Gui’an Hot Spring is SO₄·HCO₃-Na,and the rest is HCO₃·SO₄-Na.The hot spring reservoir temperature is about 66.5-130℃,and the hot water circulation depth is about 3-7 km,which is a low-medium temperature convective geothermal system.（2）An earthquake prediction model was constructed based on anomaly detection algorithms.In this study,16 hydrochemical features were selected from the original monitoring data to build data sets,and 8 anomaly detection algorithms were selected to build different earthquake prediction models,and two key parameters,anomaly detection ratio and seismic response time threshold,were set to achieve automatic prediction of the occurrence of earthquake events of magnitude 5 and above in the study area,which achieved better prediction accuracy with prediction accuracy of 76%for earthquakes occurring within 30 days while reducing the workload of manual evaluation.（3）The performance of traditional methods and machine learning algorithms in earthquake prediction are evaluated and the applicability of the algorithms are determined.The prediction performance of different algorithms was tested by using the seismic catalog data released by the China Seismic Network to validate the prediction results,taking three seismic evaluation metrics,namely,reporting accuracy,false alarm rate and R-value.The results show that the box-line plot method and Gaussian distribution model,which are widely used in current research,perform relatively poorly in prediction,the Histogram-based Outlier Score performs roughly perform roughly the same as them,the Auto-encoder algorithm and Gaussian mixture model performed slightly better,and the Isolated Forest,Local Anomaly Factor and Prophet algorithms performed best in prediction.Machine learning algorithms for earthquake prediction perform better overall than traditional methods.（4）The factors affecting the model prediction performance are analyzed to provide a basis for data pre-processing,optimization of earthquake monitoring points,selection of monitoring indicators and setting of earthquake response time thresholds.When building the seismic prediction dataset is to include the time dimension,which can effectively improve the model prediction performance.Different hot springs respond differently to earthquakes due to differences in geological conditions and hydrogeochemical characteristics,and the prediction performance of hot springs with characteristic ions varies.The earthquake response time of hot springs in the study area to seismically active areas is roughly 35-40 days...