Abstract: The room temperature of tunnel-ventilated poultry houses is easily affected by variations in the natural environment and the age of the poultry and is difficult to accurately predict online. In order to achieve accurate prediction of poultry house room temperature and improve the accuracy of temperature regulation, this paper proposes a PCA-LM-NARX method for online construction of a poultry house room temperature prediction model. The resulting NARX neural network model can accurately predict future short-term indoor temperature changes based on recent poultry house environmental data. The PCA-LM-NARX method selects the environmental variables that affect the temperature of the poultry house based on the energy balance equation of the tunnel ventilation poultry house temperature system, and uses principal component analysis (PCA) method to screen out the key factors that affect the indoor temperature of the poultry house from multiple environmental variables, construct a room temperature prediction model with a NARX neural network structure, and use the LM algorithm to optimize the model parameters. Taking into account the hysteresis characteristics of indoor temperature changes in poultry houses caused by heat conduction delay, air circulation delay, control equipment adjustment delay, evaporative cooling, etc., the optimal delay order of the NARX neural network was designed using the Bayesian information criterion. The normalization method and rolling statistical method are used to preprocess the measured data. During the modeling process, in order to adapt to poultry of different ages and environmental changes, the PCA-LM-NARX method uses the moving window method to remove old data with a large time span from the current moment in the neural network training set, and add new data with a small time span from the current moment, thereby achieving real-time updates to the training set and reducing the negative impact of old data on model prediction performance. The modeling program is divided into two parts: offline modeling and online prediction. The offline modeling part is installed on the host computer and mainly updates the parameters of the poultry house room temperature prediction model in real time based on the data in the moving window. When the LM algorithm calculation results meet the accuracy requirements, the new model parameters are transmitted to the field controller through the fieldbus network. The online prediction program and room temperature prediction model are stored in the on-site controller and used for online prediction of poultry house temperature. The field controller transmits the new training data set to the host computer through the fieldbus network for updating the prediction model parameters. Experimental results show that the poultry house room temperature prediction model established based on the PCA-LM-NARX method can predict the temperature values in the next 5 minutes, 15 minutes, and 30 minutes online, with the mean absolute error (MAE) respectively of 0.1077℃, 0.1539℃ and 0.2024℃. The shortest average modeling time of running the modeling program on the i5-12500H CPU is 0.3321s, which can achieve rapid real-time update of model parameters. The results of this study show that the PCA-LM-NARX method can be used to construct a high-precision tunnel-ventilated poultry house room temperature prediction model and achieve rapid online update of model parameters.