Predicting indoor temperature of solar green house by machine learning algorithms: A comparative analysis and a practical approach

This study focuses on a solar greenhouse located at the experimental base of Shenyang Agricultural University in Shenyang, Liaoning Province, to develop multi-step temperature prediction models based on machine learning algorithms. The research employs five algorithms: Random Forest (RF), Multiple Linear Regression (MLR), Support Vector Regression (SVR), Long Short-Term Memory Recurrent Neural Network (LSTM), and Gated Recurrent Unit (GRU) for temperature prediction. Experimental data were collected from meteorological stations inside and outside the solar greenhouse. The innovative aspect of this study lied in its systematic evaluation of temperature predictions across various time steps. Twenty-one prediction horizons, ranging from 15 minutes to 1440 minutes, were selected and the performance of the five predictive models was assessed using K-fold cross-validation for each time step. Results demonstrated that the GRU (Gated Recurrent Unit) model outperformed all other algorithms across all 21 prediction horizons, with short-term prediction (15 min) achieving an R² of 0.991 and long-term prediction (1440 min) maintaining an R² of 0.992 (as shown in Table 1). This performance significantly exceeded that of LSTM, Random Forest (RF), Support Vector Regression (SVR), and Multiple Linear Regression (MLR), with GRU reducing the root mean squared error (RMSE) by 12.3%–27.5% compared to LSTM in long-term predictions. The RF and SVR models demonstrated good performance for short-term predictions, but showed slight accuracy degradation as the prediction horizon extended. The MLR model performed adequately for short-term predictions (within 30 minutes), but exhibited poor performance for longer time steps (R² < 0.9). GRU, by virtue of its more concise gating mechanism (featuring only update gates and reset gates), not only ensured high precision but also significantly improved training efficiency compared to LSTM. This research not only compared the performance of different machine learning algorithms in solar greenhouse temperature prediction but also explored the applicability of each algorithm across various prediction horizons. The findings provide a theoretical foundation and technical support for intelligent control and precise management of solar greenhouses.