A kernel based extreme learning machine (KELM) method was proposed for predicting joint moments of lower limb rehabilitation robots to address the problem that random initialization of input weights and bias affect the accuracy of the model when predicting lower limb joint moments by extreme learning machine (ELM). The method integrated Gaussian kernel function with ELM and used genetic algorithm (GA) combined with particle swarm optimization (PSO) of genetic algorithm-particle swarm optimization (GAPSO) to optimize the parameters of KELM. Firstly, the motion data of a patient with right lower limb hemiplegia walking on a treadmill at 5 different speeds of 0.4, 0.5, 0.6, 0.7 and 0.8 m/s were collected and preprocessed. Secondly, the KELM was optimized by GAPSO to obtain the optimal regularization coefficient C and kernel function width parameter S. The output joint moments were compared with the joints calculated by inverse biomechanical analysis. Finally, the root mean square (RMSE) and correlation coefficient (P) were used to evaluate the superiority of the algorithm. The experimental results show that the average maximum root mean square error of KELM algorithm based on GAPSO (GAPSO-KELM) optimization is 14%, 18% and 28% lower than that of PSO-KELM algorithm, KELM algorithm and ELM algorithm, respectively, and the minimum P is 0.84 except for 0.8 m/s right ankle inversion, which is 0.79. The GAPSO-KELM algorithm further improves the prediction accuracy, making it a more effective algorithm support for rehabilitation treatment.