[Objective] The soil erosion under different conservation engineering measures was precisely predicted in order to provide a technical and theoretical basis for formulating appropriate conservation measures in Northwest Liaoning Province. [Methods] We used experimental plot data from 2011 to 2021 that included maximum precipitation intensity in 30 and 60 minutes (I₃₀ and I₆₀), precipitation duration (T), and precipitation (P) to construct a long short-term memory neural network model (LSTM) to predict soil erosion under three different water-and-soil conservation measures (horizontal trough, fruit tree terrace, terrace). Results from the LSTM model were compared with the results of three classical machine learning models, i.e., artificial neural networks (BP), random forest (RF), and support vector machine (SVM). [Results] ① The impacts of I₃₀, I₆₀, T, and P on soil erosion were different for the three different conservation conditions, but in general, I₃₀, I₆₀, and T had significant impacted on soil erosion. ② The normal relative mean square error (NRMSE) of the BP model under the three different water-and-soil conservation measures were all greater than 0.2. ③ Compared with the RF and SVM models, the LSTM model decreased NRMSE by 0.04~0.08, 0.02~0.08, and 0.05~0.08 under the three different water-and-soil conservation measures, respectively. ④ The LSTM model based on only two input features (I₃₀ and T) had a similar accuracy with the LSTM model based on four input features in predicting soil erosion. [Conclusion] The LSTM model was used to predict the soil erosion amount based on the maximum 30 min rainfall intensity and rainfall duration, and the prediction accuracy was higher than other traditional models. This shows that the LSTM model can be popularized and applied in the accurate simulation of soil erosion and the determination of soil and water conservation measures in similar areas.