Abstract: Security situation is the premise of network security warning.The network attacks in complex network environment bring unexpected challenges,causing the sudden network security incidents such as increasing network load and network failure happen at any time.Therefore,taking into account the uncertainty and non-linearity of network security situation time series,in order to further improve the forecast accuracy of network security situation,this paper proposed a network security situation prediction method based on NAWL-ILSTM (Nadam with Look-ahead and Improved Long Short-Term Memory).Firstly,an online updating mechanism is adopted to improve the LSTM to establish time series forecasting model,which can conduct parameter updating in real time for the received online observed data and minimize the cost function,thus solving the problem that traditional LSTM algorithm can’t use network system to transmit data online reasonably,further,optimizing the parameter updating and improving the forecast accuracy of LSTM model.Then,aiming at the problems of slow convergence speed and high training cost in the training process of neural networks,the Look-ahead technology is used to improve the updating formula of Nesterov acceleration gradient adaptive estimated momentum algorithm (Nadam) to accelerate the convergence speed of the model,and then the trai-ning speed of ILSTM prediction model can be accelerated to reduce training time and cost.The simulation experiments based on Pythonin tensorflow environment demonstrate the rationality of the LSTM prediction model based on online updating mechanism.Convergence analysis and comparison experiments show the NAWL algorithm has faster convergence speed.Finally,the comparison experiments show that the proposed model based on NAWL-ILSTM has stronger applicability and higher applicability in situation time series analysis compared with other prediction model.

Key words: Adaptive momentum estimated algorithm, Long short-term memory, Look-ahead technology, Network security situation prediction, Online observation data