This paper proposes a new deep learning method called H-LSTM (Hybrid Long and Short Term Memory) in order to improve the demand forecasting system of spare parts for ROKAF (Republic of Korea Air Force) aircraft ‘B’. The existing LSTM has been popularly utilized for forecasting stock price or energy demand since it was known to be appropriate for non-linear time series data forecast. This paper applies the H-LSTM for a demand forecast problem of aircraft spare parts, which shows irregular demand patterns. The H-LSTM that combines the existing LSTM model with time series analysis after the seasonality and trend of demand data are decomposed. Based on a preliminary analysis, the Aircraft spare parts demand pattern shows irregularity as Erratic, Lumpy items of irregular demand characteristics take relatively higher percentages. The accuracy of the new method compared with existing stochastic methods show a higher forecast accuracy than ARIMA or Holt Winters. Therefore, if it is applied for the demand forecast system of ROKAF aircraft spare parts, the H-LSTM is expected to not only improve demand forecast accuracy, but also increase aircraft availability and curtail inventory cost through decreasing unnecessary parts stocks. This paper is meaningful in that it is the very first study to offer a working-level improvement resolution in demand forecast through the LSTM, a type of deep learning, by utilizing ROKAF’s practical logistics data.