Abstract:

The traditional solution method for the incoming attacking target in water lacks the time series characteristics mining on multi-dimensional and uncertain observation data. Aiming at the time series prediction with high complexity and missing data, a method based on adaptive window interpolation and deep variable weight long-term and short-term memory network model for missing time series observation data with multiple sampling frequencies is proposed, which is compared and verified on the simulation data and public test data set. Aiming at the random missing problem caused by the inconsistency of sampling frequency of multi-source observation information, an adaptive window imputation method is introduced to adaptively adjust the imputation window according to the missing rate of observation information and reduce the error level of traditional imputation methods. The long and short-term memory network model with variable weights is introduced to time series data, and the loss weight of back propagation is adaptively adjusted according to the missing time series observation data at multiple sampling frequencies while being suitable for imputation of missing time series observation data at multiple sampling frequencies. Experiment results show that the comprehensive use of adaptive window interpolation and variable weight long-term and short-term memory network model can provide a faster and accurate decision-making reference for predicting and defending the incoming targets.

Key words: time series data, adaptive window, variable weight LSTM, target solution method