Multi-tier perching layer houses can offer hens with perches, nests, and multi-level activity spaces. But their complex spatial structure and airflow configuration can also lead to local environmental differences and short-term fluctuations. It is often required for the accurate short-term prediction of indoor environmental parameters under proactive ventilation control and environmental risk warning in cage-free laying hen production. This study aimed to develop a short-term prediction for indoor temperature, relative humidity, and ammonia concentration in a multi-tier perching layer house, according to multi-point monitoring data. Continuous environmental data were collected from one experimental multi-tier perching layer house. The indoor monitoring points were arranged in different perching areas, while one outdoor monitoring point was used to represent boundary environmental conditions. Time alignment and resampling at a 20 min interval, point-wise mean values, and spatial ranges were calculated after data cleaning to evaluate environmental differences among monitoring points. Pearson correlation analysis was used to examine the relationship among temperature, relative humidity, and ammonia concentration. A short-term prediction framework was then developed using gradient boosting regression. Historical observations, outdoor environmental information, and indoor–outdoor difference features were used as model inputs for temperature and relative humidity. A multivariable prediction model was constructed to combine ammonia historical concentration with temperature and relative humidity. The most recent 7 d data was used as an independent validation set. In addition, recursive prediction was performed to evaluate the model performance for a future 24 h horizon. Furthermore, 0.05 and 0.95 quantile regression models were established to generate 90% prediction intervals. The results showed that the indoor environmental parameters differed among monitoring points. The average temperature, relative humidity, and ammonia concentration were 18.99-20.88 ℃, 46.52%-51.98%, and 1.32-1.86 mg/m³, respectively, at different indoor monitoring points. The mean spatial ranges were 2.31℃, 11.84%, and 1.27 mg/m³, respectively, indicating that the parameters varied among different perching areas. Correlation analysis showed that temperature and relative humidity were negatively correlated with ammonia concentration, with correlation coefficients of −0.26 and −0.16, respectively. Ammonia concentration was dominated by its historical state, thermal and humidity conditions, as well as ventilation. In the independent validation set, the temperature prediction model achieved a coefficient of determination of 0.96, a root mean square error of 0.73 ℃, and a mean absolute error of 0.50 ℃, whereas those values were 0.97, 2.94%, and 1.77%, respectively, in relative humidity. In ammonia concentration, the multivariable prediction model achieved a coefficient of determination of 0.76, a root mean square error of 0.44 mg/m³, and a mean absolute error of 0.27 mg/m³. In the multivariable model, the coefficient of determination increased from 0.64 to 0.76, whereas the mean absolute error reduced from 0.37 to 0.27 mg/m³, compared with the univariate gradient boosting regression model with only ammonia historical information. As such, temperature and relative humidity features provided useful supplementary information for ammonia prediction. The 24 h recursive prediction showed that the stable prediction performance was maintained during continuous forecasting. Smooth prediction curves were produced without outstanding abnormal jumps. The 90% prediction intervals were also provided to quantify the fluctuation range of environmental parameters. The short-term prediction can be used to adjust proactive ventilation for the less ammonia risk environment in multi-tier perching layer houses.