Abstract: Subject behavior modeling of rural residential consolidation (RRC) can rely mainly on the linear utility function and prior knowledge of experts at present. These approaches have the high subjective interference, leading to hardly capture the nonlinear influence of natural, and socio-economic aspects of RRC. It is very necessary to reduce the uncertainty of simulation. In this study, an intelligent simulation model of RRC potential was constructed to couple the deep learning under the framework of agent modeling. The decision factors system and role connotations of county-level governments and farmers were constructed in the process of rural residential land consolidation in China. 37 decision indicators were selected to represent the decision-making of government agents in the aspects of urban-rural development, government finance, agricultural development, food security, location and topography, and distribution characteristics of rural settlements and cropland. 38 decision indicators were selected to represent the decision-making of farmer agents in the aspects of social economy, nature, land use and landscape pattern. As such, the light gradient boost machine and long short-term memory algorithms were utilized to construct the agent decision-making simulation modules for the governments and farmers, respectively. An agent interaction scheme was then developed using the seed region growth algorithm. Collaborative and automatic simulation was realized for the quantitative scale and spatial layout of regional consolidation potential. A case study was conducted to evaluate the effectiveness and performance of the improved model in the Guangzhou City, Guangdong Province, China. The data was collected from the county-level units in the study area. Among them, the data from county-level units in Hunan Province was also introduced to expand the sample size of the government agent module. 70% of the samples were used for the model training and 30% were used for the validation and test. The experimental results show that the government agent module was firstly achieved the determination coefficient of R²=0.9875 and the average relative error of 6.95% to predict the quantitative scale of consolidation potential. Meanwhile, the farmer agent module demonstrated an overall accuracy of 82.31% to identify the rural residential land consolidation potential. Specifically, there were a precision rate of 83.68% and a recall rate of 80.27% in the initial identification of consolidation potential. Secondly, the simulation results show that the consolidation potential of rural residential areas in Guangzhou from 2018 to 2035 was about 951.84 hm², which was mostly distributed in the areas with backward infrastructure conditions, poor living environment quality and relatively single land structure. Urban and remote mountain villages were relatively concentrated in the areas of potential distribution. In addition, the areas with spatially connected and flat terrain were more likely to be included in the consolidation scope. The simulation was basically consistent with the general framework of land use patterns. In conclusion, the strong modeling reliability and feasibility can provide the scientific and accurate data support to the land consolidation.