The meat duck industry in China has contributed over 82% of the total slaughter volume worldwide. Therefore, an optimal air temperature is often required for the comfort and survival of meat ducks during breeding. However, the air temperature is susceptible to multiple factors, such as relative humidity and illumination intensity. It is a high demand to timely and accurately predict air temperature for the high-density healthy breeding. However, conventional temperature prediction has been limited to low accuracy, robustness, and generalization. In this study, a hybrid 1DCNN-BiLSTM-DQN model was proposed to integrate with a one-dimensional convolutional neural network (1DCNN), a bidirectional long short-term memory network (BiLSTM), and a deep Q-network (DQN). Duck-house temperature was accurately predicted after model construction. The temperature time-series signal was also decomposed into high- and low-frequency components via the discrete Fourier transform (DFT). Given that the high-frequency component represented short-term fluctuations, the 1DCNN was used to extract local features from the high-frequency component; whereas the low-frequency component represented long-term fluctuations, the BiLSTM was used to extract long-sequence dependency features from the low-frequency component. Subsequently, the two sets of features were fused using a concatenation model. And finally, the temperature prediction value was obtained after the mapping of a fully connected layer. Furthermore, the DQN algorithm was introduced to construct an agent for iterative optimization of the hyperparameters. An adaptive mechanism was optimized with an 8-dimensional state space and a 12-dimensional action space, enabling dynamic optimization of key hyperparameters, such as the learning rate, number of hidden layers, dropout rate, and network architecture. Thereby, the prediction robustness of the model was improved under scenarios of seasonal transitions and extreme weather. The indoor and outdoor temperature data of net-raised duck houses were collected in Gaoyou City, Yangzhou City, Jiangsu Province, from March 22, 2025, to March 22, 2026. The results demonstrated that the 1DCNN-BiLSTM-DQN model achieved a coefficient of determination (R²) of 0.993 with an optimal input step size of 48. The MAE and RMSE were superior to the conventional models, such as the Temporal Convolutional Network and Transformer. Specifically, the 1DCNN-BiLSTM-DQN model exhibited the following improvements under different weather conditions: On cloudy days, the MAE and RMSE decreased from 0.29 °C to 0.18 °C, and from 0.36 °C to 0.23 °C, whereas the R² increased from 0.77 to 0.91; On sunny days, the MAE and RMSE decreased from 0.48 °C to 0.23 °C, and from 0.59 °C to 0.31 °C, whereas the R² increased from 0.79 to 0.94; On rainy days, the MAE and RMSE decreased from 0.54 °C to 0.33 °C, and from 0.70 °C to 0.46 °C, whereas the R² increased from 0.72 to 0.89. The combined 1DCNN-BiLSTM prediction model achieved better prediction performance compared with the conventional models, such as BiLSTM, 1DCNN, TCN, and Transformer. In summary, the 1DCNN-BiLSTM-DQN model can be expected to predict the temperature in duck houses. The findings can also provide data support for early environmental regulation, thereby reducing the risk of environmental stress in meat ducks.