Battery energy storage (BES) participation in the grid ancillary services markets is increasing rapidly in recent years. To facilitate optimal participation, the need for accurate BES state-of-charge (SOC) forecasting is indispensable. In grid ancillary services, the development of SOC forecasting models should deal with uncertainties and corresponding stochastic processes that determine the BES SOC periodically. Several traditional and state-of-the-art machine learning (ML) techniques, ranging from decision-tree to deep learning methods, were used to solve this problem. However, developing a multi-step SOC forecasting model remains a challenge in this subject that is essential for optimal BES economic dispatch and unit commitment. Taking advantage of the Long short-term memory (LSTM) deep learning and its variants techniques which are proven to be a robust method for predicting sequentially dependent data in the time-series domain, this paper proposes LSTM-based multi-step SOC forecasting for BES operating in frequency regulation. Various developed models, i.e., Vanilla-LSTM, Vanilla-Gated Recurrent Units (GRU), Bidirectional-LSTM (Bi-LSTM), and Bi-GRU, are evaluated using real-world datasets. The evaluation results show that the developed models outperform the existing methods in terms of root mean square error (RMSE) and mean absolute error (MAE) evaluation metrics.