Research On Accuracy Optimization Method Based On Large-scale Light Curve Forecasting Model

In the field of astronomy,major astronomical events are often accompanied by sudden changes in the brightness value of light curve.Therefore,the analysis of light curve can better capture major astronomical events.The main goal of this paper is to realize the early warning of abnormal astronomical events and improve the real-time and comprehensive processing of astronomical observation problems by predicting and analyzing the light curve of celestial bodies.Furthermore,we should promote the discovery of supernovae,detect a series of important astronomical problems such as gravitational microlensing and gamma ray burst timely and accurately,and promote the development of astronomical big data.At present,the prediction task of the light curve sequence is still in the stage of statistical method.In the face of time sampling frequency of 15 s,there are hundreds of astronomical data generated every moment,which cannot effectively guarantee the prediction accuracy and time efficiency.At the same time,because the celestial data under high sampling frequency will introduce a lot of interference noise,which is not conducive to analysis and modeling.In terms of model itself,traditional statistics and machine learning technology cannot capture the long-term dependence on the time dimension.Finally,in the aspect of anomaly detection of light curve,we need to set the threshold artificially,which is seriously affected by subjectivity,and lack of adaptive detection methods,so anomaly detection cannot achieve high robustness.In order to solve the problems of large amount of data,high sampling frequency,insufficient prediction accuracy and low robustness of anomaly detection in analysis of time series of optical curves,this paper proposes a prediction model of optical curves based on recurrent neural network and a anomaly detection method based on convolution neural network.Light curve studied is a graph that astronomically represents the brightness change of celestial bodies relative to time.The data used in the model is the image taken by mini-GWAC of gamma ray burst exploration astronomical satellite cooperated by China and France.Light curve data are obtained after point source extraction and cross validation.The method of data extraction is to connect the multi sky data obtained from the same star in order to form a data stream for prediction and analysis.The main problems and contents of study include prediction of light curve and anomaly detection of light curve,the details are as follows:(1)The prediction method of light curve based on LSTM neural network of time series decomposition: it mainly optimizes the accuracy of the existing prediction model,achieves the fitting of light curve of the model to the specific sky area,and makes the prediction more accurate and fast.The main prediction model is LSTM prediction model based on EMD decomposition.The time-frequency analysis method based on EMD can be applied to any type of signal decomposition.It has obvious advantages and high signal-to-noise ratio in processing non-stationary and non-linear data.LSTM is the most perfect algorithm for time series data modeling,which can retain the information of the previous time by loop feedback connection,so that the whole model can capture the long-term sequence information.At the same time,ARIMA,MLP,SVR,RNN and GRU models are built for comparative experiments to verify the prediction accuracy and effect of the proposed prediction model.(2)On this basis,the anomaly detection task of light curve is realized.In this paper,an anomaly detection method based on spectral residual data transformation and convolution neural network is proposed.CNN is directly applied to the output data of SR model,and CNN is used to learn a discrimination rule to alleviate the single threshold problem of the original SR method.In anomaly detection,our goal is to detect astronomical anomalies in a short time scale and provide early warning.Here we focus on the anomaly detection task of gravitational microlensing.The method of deep learning is used to adapt to non-static data,so that the anomaly detection model no longer clearly requires and defines the original data distribution,which shows excellent accuracy and robustness compared with the traditional statistical methods.