As China’s poultry industry is rapidly moving towards standardization and large-scale development, high-yield breeds in large-scale farms are particularly sensitive to environmental changes. Temperature and humidity are core factors affecting poultry health and production efficiency: excessively high or low temperatures can slow down growth and reduce feed conversion rates, while abnormal humidity may increase the risk of pathogen transmission. Currently, most environmental control systems in poultry houses rely on comparing real-time sensor data with preset thresholds to activate equipment. However, there is a delay in equipment response, making it difficult to cope with dynamic environmental changes. How to accurately predict future temperature and humidity and optimize control strategies in advance has become a key issue for improving poultry farming efficiency.
A research team led by Professor Guanghui Teng from the College of Water Resources and Civil Engineering, China Agricultural University, has developed a multi-step prediction model for temperature and humidity called “GFF-transformer”. Innovatively integrating a “gated feature fusion (GFF) module” with a transformer architecture, this model achieves high-precision prediction of temperature and humidity in poultry houses for the next 6, 12, 18, or 24 hours by extracting multi-scale temporal features and capturing long-term dependencies. The relevant research has been published in Frontiers of Agricultural Science and Engineering (DOI: 10.15302/J-FASE-2025603).
Traditional prediction models such as LSTM and GRU can handle time-series data but have limitations in capturing global information and computational efficiency. Although the transformer model excels in long-sequence analysis, it struggles to effectively extract local key features. The GFF-transformer model resolves this contradiction through a parallel processing mechanism: the GFF module uses three groups of 1D convolutional layers with different group numbers (1, 2, 4) to capture short-term, medium-term, and long-term environmental features respectively, and selects key information via a gating mechanism similar to that of LSTM; the transformer encoder models complex correlations between variables through a self-attention mechanism. The features extracted by the two components are fused, and the prediction results are output through a fully connected layer.
Experimental data show that within the temperature range of 20.1–31.5 ℃, the model achieves a coefficient of determination (R²) of 0.88–0.92, a mean absolute error (MAE) of 0.48–0.62 ℃, and a root mean square error (RMSE) of 0.68–0.85 ℃ for 6–24-hour predictions. Within the humidity range of 18%–97%, the R² reaches 0.86–0.94, the MAE is 2.9%–4.7%, and the RMSE is 4.3%–6.4%. Compared with LSTM, GRU, and the traditional transformer model, the GFF-transformer model increases the R² of temperature prediction by 0.01–0.08 and reduces the MAE by 0.05–0.24 ℃; for humidity prediction, it raises the R² by 0.03–0.08 and lowers the MAE by 0.81%–1.43%, demonstrating stable performance especially in 24-hour long-term predictions.
Studies indicate that the model’s advantage lies in its ability to simultaneously process multi-source environmental data (including temperature, humidity, CO₂, and static air pressure) and improve prediction robustness through interactions between variables. For instance, enhanced ventilation in summer leads to a decrease in CO₂ concentration—this correlation feature is effectively captured by the model, thereby optimizing the accuracy of temperature and humidity prediction. In addition, the model takes 504 seconds (approximately 8.4 minutes) for training and 0.25 seconds for a single prediction, which can meet the real-time requirements of environmental control in poultry houses.
This research provides a practical tool for the intelligent management of poultry house environments. By accurately predicting temperature and humidity 6–24 hours in advance, farms can dynamically adjust the operation strategies of ventilation and temperature control equipment, an
DOI: 10.15302/J-FASE-2025603