Abstract: Accurate short-term prediction of indoor air temperature can often dominate the intelligent and proactive environmental control in modern laying hen houses. However, the thermal environment inside poultry facilities is generally governed by complex dynamics with pronounced nonlinearity, non-stationarity, and multi-scale temporal characteristics. However, conventional statistical models and standard deep learning have confined to the relatively slow-varying or rapid fluctuations during operations and environmental control equipment, such as the diurnal and seasonal cycles. The significant challenges are also the struggle to simultaneously capture the long- and short-term variations in the agricultural environmental time series. In this study, a hybrid temperature prediction framework was proposed to integrate the Empirical Mode Decomposition (EMD), Long Short-Term Memory (LSTM) networks, and an attention mechanism, termed the EMD-LSTM-Attention model. The representation and deep learning of the multi-scale temporal features were enhanced in the temperature data of the laying hen house. Specifically, the EMD was first employed as an adaptive signal decomposition. The original non-stationary temperature time series were decomposed into a finite number of intrinsic mode functions (IMFs). Each oscillatory component corresponded to different characteristic time scales. These IMFs were then treated as multi-channel inputs and fed into an LSTM-based prediction network equipped with an attention mechanism, learn temporal dependencies in each decomposed component. The greater importance was dynamically allocated to more informative historical-time steps during prediction. The continuous indoor temperature data were collected from a commercial laying hen house over an entire year, with the measurements at 10-minute intervals. The model robustness was examined under different seasonal conditions. The dataset was organized into five independent but temporally continuous subsets: full-year, spring, summer, autumn, and winter. A chronological split was applied in each dataset, thus allocating 60% of the data for training, 20% for validation, and 20% for testing. All models were trained under identical experimental settings using the Adam optimizer. An early stopping strategy was used to reduce the overfitting. The performance was assessed using three widely adopted evaluation metrics: mean absolute percentage error (MAPE), root mean squared error (RMSE), and the coefficient of determination (R²). The EMD-LSTM-Attention model was evaluated with multiple benchmarks, including the conventional autoregressive (AR) models and several commonly-used deep learning architectures, such as the RNN, GRU, BiGRU, BiLSTM, and Transformer. In addition, ablation experiments were conducted to quantify the contributions of the EMD and attention modules. The results demonstrate that the EMD-LSTM-Attention model consistently outperformed the benchmark models over all datasets. The better performance was achieved in a full-year MAPE of 1.43%, an RMSE of 0.291 °C, and an R² of 0.952 under the default experimental configuration, with an input sequence length of 4 h and a prediction interval of 10 min. The superior prediction accuracy was also compared with the baseline. Furthermore, the relatively low prediction errors were obtained in the spring, summer, autumn, and winter datasets, indicating the stable performance under varying seasonal thermal conditions. A parameter sensitivity analysis was conducted to investigate the influence of key input settings on the model performance. The lowest prediction error was achieved in the range of input sequence lengths using a 1.0 h historical input window. The full-year MAPE decreased from 1.43% (with a 4.0 h input window) to 0.76% under this configuration. Additionally, the impact of the prediction horizon was examined to extend the forecast length from 0.5 to 2.5 h. The model errors also exhibited an overall upward trend, as the prediction horizon increased; Nevertheless, the prediction accuracy remained at a relatively high level in the tested range. Overall, the findings. The signal decomposition with deep learning models substantially enhanced the accuracy and robustness of the short-term temperature prediction in the laying hen houses. The EMD-LSTM-Attention framework can provide a viable data-driven solution to capture complex thermal dynamics. The practical support can also be offered for the predictive and intelligent environmental control in poultry production. The additional environmental variables can be incorporated to further improve the model generalization and applicability over multiple housing facilities.