Long-term positioning experiment was used to set up four sorts of tillage measures (traditional tillage, no-tillage, traditional tillage straw mulching and no-tillage with straw cover) for the rotation sequence of spring wheat/pea and pea/spring wheat. The monthly precipitation, monthly average radiation, monthly average evaporation, and monthly crop water consumption were used as input factors, and monthly average soil water content was used as an output to establish a prediction model of soil water content based on long short-term memory (LSTM) neural network, and the validity of the model was evaluated to analyze the differences of soil water content effect of different conservation tillage measures in the Loess Plateau gully region of central Gansu, then the model was applied to simulate the dynamics of the soil water content under four tillage measures in the 0~200cm soil layer. The results demonstrated that the soil water model based on LSTM neural network had good applicability for predicting soil water content under conservation tillage in the Loess Plateau gully region of central Gansu, the average root mean square error, mean relatively error and determination coefficient of the model were 2.29%, 6.79% and 0.82, respectively. In pea/spring wheat rotation sequence, the soil water content of four treatments was increased by 1.49%, 1.61%, 1.69% and 1.76%, respectively, compared with spring wheat/pea, the descending order of soil water content of the four tillage measures was as follows: no-tillage with straw cover, no-tillage, traditional tillage straw mulching, and traditional tillage in 0~200cm soil layer, the average soil water content of no-tillage with straw cover was increased by 1.27%, 1.75% and 2.81%, respectively, compared with no-tillage, traditional tillage straw mulching, and traditional tillage. The no-tillage with straw cover had the most significant effect on soil water content in 0~30cm soil layer, the average soil water content of no-tillage with straw cover was increased by 1.60%, 2.63% and 4.18%, respectively, compared with no-tillage, traditional tillage straw mulching, and traditional tillage. Soil water content of four tillage measures was changed with seasons, but the soil water content effect of no-tillage with straw cover was better than that of the other three tillage measures, the effects of water storage and soil moisture conservation were more significant during pre-crop growth. The LSTM neural network model achieved good simulation results on the soil water content in the Loess Plateau gully region of central Gansu, the soil water content of four tillage measures in the pea/spring wheat rotation sequence was relatively higher, no-tillage with straw cover was beneficial to improve the soil water content of farmland in the study area the under four tillage measures, and the most suitable one was conservation tillage measure in the Loess Plateau gully region of central Gansu.