Human Trajectory Forecasting with RNN-Based Hybrid Models: LSTM, GRU, and SimpleRNN Combinations

For 5G aerial base station (UAV-BS) networks to have a smooth and dependable handover, human trajectory prediction is crucial. Although the majority of research focuses on individual architectures rather than hybrid approaches, deep learning models such as SimpleRNN, GRU and LSTM demonstrated potential in modeling sequential data. There are currently few comparative studies of hybrid designs, especially when dropout regularization is used. The SimpleRNN-Dropout-LSTM-Dropout model produced the best results out of all of them after 50 epochs of training with Tanh activation, Adam optimizer, learning rate of 0.001, 64 hidden units, and batch size of 32. With little training and validation loss (0.0007 each), it recorded the lowest errors across all metrics: MSE (0.0003), MAE (0.0146), ADE (0.0207), FDE (0.0434), RMSE (0.0168), and MAPE (0.0157). These results demonstrate how well hybrid deep recurrent networks - in particular, SimpleRNN-LSTM combinations - perform in accurately predicting short-term human trajectories.