Comparative Analysis Of Machine Learning Models For Greenhouse Microclimate Prediction

Abstract

The research assesses the effectiveness of these models as Bi-LSTM, ANN, GBM, and RF in predicting microclimate factors like temperature, humidity, and CO2 levels. It also highlights the constraints associated with employing machine learning models for greenhouse microclimate prediction and suggests potential areas for future investigation. The findings indicate that both ensemble techniques (Gradient Boosting Machine and Random Forest) and deep learning frameworks (ANN and BI-LSTM) performed well during the assessment. While both ensemble methods exhibited impressive results, Gradient Boosting Machine (GBM) slightly surpassed Random Forest (RF) across various evaluation criteria. GBM attained a notable R-squared value of 0.9998, signifying its robust fit and capacity to elucidate data variability, in addition to a Root Mean Squared Error (RMSE) of 0.0079 and Mean Absolute Error (MAE) of 0.0001. RF demonstrated similar outcomes, with an R-squared value of 0.9999. Conversely, ANN outperformed BI-LSTM in terms of R-squared values and MAE, displaying an R-squared value of 0.999999 and a MAE of 0.0079. An analysis of the sensitivity of the ANN model revealed that altering the average indoor relative humidity in the first sensor had the greatest impact on the prediction outcome among other variables. Assessing and ranking the importance of each feature used in training the RF and GBM models indicated that the average relative humidity in the second sensor held the highest significance, with any modification to it likely to notably influence the prediction outcome. These results support the notion that machine learning algorithms serve as effective predictive tools, offering valuable insights for enhancing greenhouse operations. Future research should focus on practical implications and real-world applications, particularly in optimizing hyperparameters.