Optimizing Greenhouse Management with Accurate Temperature Prediction

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Aug 11 2023

In an article published in the journal Agriculture, researchers explored accurate temperature prediction for greenhouse management. To enhance crop yields and mitigate losses, optimal control strategies and minimizing energy waste were considered. They compared multiple linear regression (MLR), Radial Basis Function (RBF), and Support Vector Machine (SVM) models for forecasting greenhouse air temperature using external inputs.

The RBF model with the Levenberg–Marquardt (LM) learning algorithm stood out, achieving the lowest error and highest coefficient of determination (R2) value. With a Root Mean Square Error (RMSE) of 1.32°C, the Mean Absolute Percentage Error (MAPE) of 3.23%, and R2 of 0.931, the RBF model reliably predicted greenhouse temperatures within the next two hours. Implementing this model could optimize time management and energy use, increasing overall greenhouse efficiency.

Related work

Past studies have highlighted the pivotal role of greenhouse air temperature in crop yield, emphasizing the need for efficient heating and cooling methods. Strategies involving shading, ventilation, and alternative energy use have been advocated to uphold optimal conditions, leading to better yields. Solar radiation and crop interaction impact the energy balance, while real-time temperature predictions aid proactive measures.

Researchers have explored Artificial Intelligence (AI) techniques, such as Artificial Neural Networks (ANN), Adaptive Neuro-Fuzzy Inference System (ANFIS), Least Squares Support Vector Machine (LSSVM), Partial Least Squares Regression (PLSR), and SVM, for temperature forecasting, yielding improved crop yields. In the Iranian city of Jiroft, studies have addressed the temperature forecasting gaps, benefiting greenhouse management and enhancing crop yields. However, a gap remains in temperature forecasting for upcoming hours. These studies have aimed to predict indoor greenhouse temperatures using MLR, RBF, and SVM models, select the optimal option for the implementation, and assess its performance in foreseeing future temperatures, especially during temperature-sensitive crop seasons. These endeavors contribute to improved greenhouse management, fostering higher crop yields and minimizing economic impact in Jiroft City.

Proposed method

The present study harnessed a range of predictive models, including MLR, ANN, RBF, and SVM, to forecast greenhouse temperature. MLR integrated external parameters like outside air temperature, relative humidity, wind speed, and solar radiation as inputs. The ANN model comprised input, hidden, and output layers, each with specific functions, while RBF excelled in capturing complex nonlinear relationships. RBF's hidden layer neurons utilized nonlinear activation functions, and the SVM model proved adept at solving quadratic optimization problems. Data normalization and K-fold cross-validation were applied for improved performance reliability.

In assessing model accuracy, the study employed critical evaluation metrics such as RMSE, MAPE, and R2. The goal was to gauge how closely the predictions aligned with actual greenhouse temperatures. Through this comprehensive approach, the research contributes to the advancement of greenhouse management practices, facilitating better prediction of temperature conditions for optimal crop growth and yield.

Experimental results

Data was collected from a greenhouse in Jiroft, Iran, which is located at 28°40′ latitude and 57°44′ longitude. The greenhouse employed natural ventilation, cooling, and heating systems as needed. Sensors captured air temperature, humidity, wind speed, and solar radiation data. AM2303 sensors recorded temperature and humidity, the TES132 solar power meter collected radiation data, and the DT186 anemometer measured wind speed. Data was collected from 8 am to 4 pm over a month in 2020 during cucumber cultivation.

A comprehensive sensitivity analysis was conducted to determine effective inputs and optimal training data sets. Correlation coefficients indicated intricate relationships among inputs, excluding wind speed. Solar radiation, outside air temperature, and relative humidity showed significant correlations. The analysis identified the best input combination for the MLR model, which was then used as a basis for the RBF and SVM models. The wind speed displayed less relevance, and the selected inputs included solar radiation, outside air temperature, and relative humidity for all models. The dataset was divided into different MLR and RBF model strategies, with the 80-20% split yielding the best results. The quadratic MLR model demonstrated superior performance.

In the RBF model, optimizing the spread parameter and neuron count significantly impacted the prediction error. An optimal structure of 3-21-1 neurons with the LM algorithm was selected, achieving high accuracy. The SVM model, utilizing the RBF kernel function, outperformed other functions in temperature prediction. The models were evaluated by comparing actual and predicted values, demonstrating strong accuracy in the RBF model with a determination coefficient of 0.93. Additionally, the ability of the RBF model to predict greenhouse temperature was validated for different time intervals, with higher accuracy for shorter intervals, highlighting its potential for effective greenhouse temperature management.

Conclusion

This study aimed to predict indoor greenhouse temperatures through MLR, RBF, and SVM models, yielding the following key findings. A sensitivity analysis using the MLR model revealed wind speed's minimal impact on temperature prediction, leading to its exclusion from all models. Model comparison highlighted the RBF model's suitability for temperature prediction in conventional greenhouses, showcasing RMSE values of 1.3 °C during training and 1.38 °C during testing. The RBF model accurately predicted greenhouse temperature for the next two hours.

Additionally, the study emphasized the potential of ANN models in advancing smart greenhouses for improved time management, reduced plant stress, energy efficiency, and enhanced crop yields in Jiroft City's agricultural landscape. For model robustness, the research advocated collecting long-term data and accounting for climate change effects on weather conditions to develop more accurate models that optimize resource usage and enhance plant growth in smart greenhouses.