Research On Prediction And Early Warning Of GWAC Light Curve Based On LSTM Neural Network

Astronomical big data is an important part of the research in the field of science.The study of astronomical phenomenon is the necessary foundation for us to understand the universe and solve the problems related to astronomy.The study of light curve can find transient astronomical phenomena.GWAC is the importrant ground-based equipment of SVOM project co-operated by China and French,which provides important point-source image information.As for researching on light curve field,it is very important to select appropriate method to monitor and predict the real time luminance data.In this paper,we firstly summarize the background and significance of our subject.Then we analyse the significance of LSTM neural network for the subject based on the characteristic of light curve.And the essence of light curve is defined.We also summarize the research production of light curve and related anomaly detection methods in various fields from both home and abroad.The related technology theory is introduced,which is the basic and key technique in this paper.Then the light curve database used in this paper is introduced.It is the process of generating the light curve from the point source image to the luminance curve.The characteristic values of the point source information are extracted in the database,including unique ID,time stamp and luminance value of each star.The selected data set is analyzed to get a better understand.According to the extracted feature values,we pre-processed them as the usage data of our models.Using the training of LSTM neural network model line data,the process of model structure and parameter selection is introduced.The performance evaluation of the model is carried out,and an anomaly detection method based on prediction error is proposed based on the error of model prediction and real value.After that,the offline learning and early warning algorithm is improved,so that an online learning early warning algorithm based on LSTM neural network is proposed.Compared with off-line algorithm,the online learning process of the model updates the weight with the new observation data.The prediction error caused by the data acquisition of real time curve data is deviated from the normal value,which leads to the prediction deviation of the normal luminance data.The update of the weight of the suspected outliers is different from the normal point algorithm,so that the prediction results after the outliers can be updated to the normal state quickly.Based on the optimized prediction results,anomaly detection based on statistical method is used.Experiments show that it has better results compared with off-line algorithm.