Multi-step prediction of greenhouse temperature and humidity based on temporal position attention LSTM

Changes in the indoor environment of greenhouses significantly impact crop production. These microclimate changes, influenced by various environmental factors, can cause crop yield and quality fluctuations. Accurate prediction of changes in these environmental factors is essential to preventing adverse impacts. However, greenhouse interiors are multivariable coupled systems with nonlinear, hysteretic, and oscillatory behavior, posing challenges for short-term predictions. Addressing this, this study proposes a novel short-term multi-step prediction model for greenhouse microclimate using Temporal Position Attention Long Short-Term Memory (TPA-LSTM) based on position encoding and time sequence. The model uses true past environmental data, including indoor and outdoor air temperature and humidity, sunlight intensity, and atmospheric pressure, employing a sliding window strategy with different lengths (30 min, 1 h, 2 h, 4 h) to predict future climate changes accurately in air temperature and humidity from 30 min to 240 min. Compared with the well-known models Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), the proposed model achieves higher credibility and lower errors. In the single-step prediction, the model with a 2 h sliding window yielded the best results, with \(\text {R}^{2}\) of 0.992 and Root Mean Square Error (RMSE) of 0.715 for air temperature and \(\text {R}^{2}\) of 0.983 and RMSE of 2.775 for air humidity, respectively. The model achieved better results for multi-step predictions of 2, 4, and 8 steps than other well-known models. The results demonstrate that this model has superior short-term prediction ability for greenhouse microclimate, significantly reducing prediction errors, enhancing environmental monitoring and early warning, and promoting intelligent greenhouse development and facility agriculture.