Greenhouse temperature prediction model based on ISGA-AMSCNN-DD

Accurate short-term temperature prediction in greenhouses is crucial for enabling advanced control strategies, such as model predictive control. However, the nonlinear and time-varying interactions between external weather conditions, crop transpiration, and environmental regulation devices pose significant challenges to forecasting accuracy. To address these challenges, we propose a novel hybrid prediction model, ISGA-AMSCNN-DD, which integrates the Improved Snow Goose Algorithm (ISGA), an attention-enhanced Multi-Scale Convolutional Neural Network (AMSCNN), and a dendritic network architecture (DD). Specifically, ISGA uses a classification-guided optimization strategy to improve global search capability and effectively fine-tune model parameters. The AMSCNN module enhances feature extraction by combining multi-scale convolution structures embedded with attention mechanisms to capture complex temporal and spatial dependencies. The DD architecture, in contrast, focuses on enhancing the model's generalization ability by analysing key relationships within the data. We validate the proposed model using real-world greenhouse temperature data collected from vegetable greenhouses in Fengyun Village, Nanchuan District, Chongqing, China. Experimental results demonstrate that ISGA-AMSCNN-DD achieves superior performance, with an R² of 0.9796, PBIAS of 0.1218%, NSE of 0.9849, RMSE of 0.6232°C, and MAPE of 2.8623%, confirming its accuracy and reliability. These findings validate the effectiveness and robustness of our approach, providing a strong foundation for intelligent greenhouse management and temperature control.