Traditional methods for collecting ground reaction forces (GRFs) mainly use lab force plates. Previous research broke this pattern by predicting GRFs with deep learning and data from IMUs like joint acceleration. Joint angle, as a geometric, is easier to collect than acceleration outdoors with cameras. LSTM is one of the deep learning models that have shown good performance in biomechanical studies. xLSTM, as an optimized version of LSTM, has not been used in biomechanical studies and no research has predicted GRFs during running solely using lower limb joint angles. This study collected lower-limb joint angle and vertical ground reaction force data at five speeds from 12 healthy male runners with Xsens sensors. Datasets including three joints and three planes were set as the inputs of four deep learning models for vertical-GRF prediction. CNN-xLSTM consistently performed best in the four deep learning models when different datasets were input (R² = 0.909 ± 0.064, MAPE = 2.18 ± 0.09, rMSE = 0.061 ± 0.008), and the performance was at a relatively high level at the five speeds. The current findings may contribute to a new GRF measurement and provide a reference for future real-time motion detection and sport injury prediction.