Fuzzy adaptive-normalized deep encoder-decoder network: Medium and long-term predictor of temperature and humidity in smart greenhouses

Abstract

To achieve accurate medium- and long-term predictions of greenhouse temperature and humidity, this study proposes a fuzzy adaptive normalized encoding and decoding network. Initially, a fuzzy preprocessing method is applied to temperature, humidity, and weather prediction data from the Internet of Things, reducing the impact of volatile fluctuations on prediction accuracy. Additionally, a fusion of the fuzzy module with an adaptive normalization technique is introduced to prevent issues like neural saturation and gradient vanishing in deep networks. The framework integrates a Codec model with the adaptive normalization module, incorporating attention normalization to enhance long-term data dependencies and improve predictive performance. The experimental dataset includes hourly temperature and humidity readings collected throughout 2020 from a smart greenhouse in Weifang, Shandong, China, which serves as the training data. Evaluation conducted on data from January 1 to February 15, 2021, shows that the proposed model outperformed all comparative approaches, particularly excelling in long-term greenhouse environment predictions. In terms of RMSE, the AN-FSANC model improved temperature prediction accuracy by 30.8 %, 24.6 %, 29.6 %, 35.3 %, 35.1 %, 16.7 %, 25.3 %, and 24.9 % compared to LSTM, GRU, BiLSTM, BiGRU, Attention-LSTM, Attention-GRU, SAN+LSTM, and AN-SANC, respectively. The model also achieved the highest performance in temperature and humidity predictions, as demonstrated by R2 and MAPE metrics, providing valuable support for agricultural applications.