Abstract: Underwater sensor network enables humans to detect and develop marine resources， but the data collected by underwater sensor nodes loses their value when accurate localization information is lost. For many underwater localization algorithms that have been widely used are still difficult to achieve accurate ranging, therefore their localization accuracy is too low and unsatisfactory. This paper proposes an underwater localization algorithm based on Long Short-Term Memory modified ranging value to improve localization accuracy. The algorithm uses a variant model of Recurrent Neural Network, namely Long Short-Term Memory (LSTM), to improve the time difference of arrival ranging algorithm. LSTM is trained by the historical data of the marine environment and the ranging value. It can efficiently and accurately predict the current ranging correction value, so as to achieve the effect of optimizing the ranging error. And the effective combination of the above two is utilized to further improve the multilateral localization algorithm to achieve precise positioning of unknown underwater nodes. Finally, the simulation experiment and algorithm comparison prove that the algorithm proposed in this paper does have high localization accuracy and feasibility.

Key words: underwater localization algorithm, range correction, Long Short-Term Memory(LSTM), underwater wireless sensor networks