Abstract: The Dst index is one of the widely used hourly geomagnetic indices to reflect geomagnetic storm processes, and forecasting the Dst index constitutes a primary concern in modern space weather studies. This study leverages Long Short-Term Memory (LSTM) neural network methodology alongside solar wind parameters and Dst index data spanning from 2008 to 2022 to construct a predictive model for the Dst index. Two models are established: the LSTM model, modeling the entire temporal domain, and the Storm model, modeling solely data from storm periods. Employing the LSTM model for rolling forecasts of Dst index during 2001 to 2002 yields a correlation coefficient exceeding 0.94 and a root mean square error within 11 nT for forecasts ranging from 1 to 6 hours in advance. The Storm model effectively addresses the issue of pronounced forecast errors during storm periods, particularly during the main phase of intense storms (Dst < –100 nT), showcasing improved forecast accuracy. Forecasting experiments conducted on 23 strong storm events occurring during 2001―2002 demonstrate an enhancement in the correlation coefficient for forecasts made 6 hours in advance during storm periods, increasing from 0.902 with the LSTM model to 0.948 with the Storm model. Integration of both forecasting models into the LSTM-Storm model yields correlation coefficients above 0.95 and root mean square errors within 9 nT for Dst index forecasts, presenting a marked improvement in forecasting accuracy compared to the standalone LSTM model.