Improvement of the Greenhouse Temperature Analysis Model for Explanatory and Adaptability

The greenhouse is a typical microclimate regulation system, where the indoor environment is influenced by numerous external variables and is prone to significant overall fluctuations. Under such circumstances, conventional methods such as fixed temperature thresholds and linear control are inadequate in addressing the energy wastage issues caused by excessive temperature adjustments. This paper investigates a multiparameter analysis method for sunlit greenhouses centered on historical data mining, aiming to provide reasonable temperature predictions for sunlit greenhouses through multiparameter analysis while enhancing model interpretability, thereby laying a foundation for precise regulation of the greenhouse environment. Unlike traditional analysis networks that rely on fixed parameters, this study proposes a framework that combines fuzzy networks with parameter optimization. It utilizes a priori superior architectures to implement fuzzy settings for the CNN network, employs KAN modules to improve interpretability, and conducts global parameter optimization. The training costs are reduced by 9% based on pre-training, while the accuracy increases by 11%. Furthermore, as the number of parameters continues to rise, the training accuracy consistently improves. The model is also capable of fine-tuning optimization under varying weather conditions, thus mitigating the impact of significant fluctuations in the internal environment of the sunlit greenhouse on model accuracy.