Abstract: To explore the applicability of Recurrent Neural Network (RNN) to simulate the long-term perceived loudness and harmoniousness of soundscape (PLS, PHS) in urban parks, this research uses the Elman and NARX neural networks with temporal memory and delay functions for separate validation. It takes objective parameters of soundscape and lightscape as inputs, and PLS and PHS as outputs to train and simulate the models of Elman and NARX neural network. The results show that: 1) There is correlation between PLS, PHS, LAeq, background sound (L95) and direct light (TL); 2) The Elman neural network has better long-term simulation effect on PLS and NARX neural network has better long-term simulation effect on PHS; 3) From the contribution rate of each parameter in the input layer, LAeq, foreground sound (L5), psychoacoustic parameter loudness (LO) and diffuse light (EL) simultaneously show a higher contribution in the two RNN models. The results indicate that the RNN is suitable for the long-term perceived soundscape in urban parks, providing an effective method and reference for the optimal design of soundscape in urban parks.