Greenhouse Climate Forecasting: an RNN-LSTM approach

Abstract

Accurate climate forecasting in greenhouses is crucial for efficient resource management and optimizing crop yields. Traditional forecasting methods often struggle due to the complexity of these ecosystems influenced by variables like temperature, humidity, light, and CO2 levels. Advanced machine learning techniques, particularly RNNs and LSTMs, excel in managing this complexity by handling sequential data and long-term dependencies. However, each technique has its limitations. To address these limitations, a hybrid RNN and LSTM model was introduced, improving forecasting accuracy and reliability. Trained on the Jena weather time series dataset and evaluated on the Algeria Chapter Green dataset, the model, featuring two layers of RNN and LSTM with 64 cells each, achieved high accuracy with low MAE and MSE values, notably for air pressure (0.73 mbar, 0.12 mbar), temperature (0.76°C, 0.12°C), and air density (3.59 g/m3, 0.58 g/m3). The coefficients of determination (R2) for these parameters were 0.9723, 0.9886, and 0.9899, respectively. These findings demonstrate the potential of advanced neural networks in improving greenhouse management and agricultural productivity.