Federated Learning for Greenhouse Temperature Prediction Using LSTM: A Privacy-Preserving Approach With Gradient-Based Aggregation

This chapter presents a federated learning-based approach to greenhouse temperature prediction using a Long Short-Term Memory (LSTM) model, addressing the challenges of data privacy and computational constraints in a distributed environment. The method incorporates an ensemble function for federated learning based on the gradients of the Mean Square Error (MSE) to optimize model performance while ensuring data privacy. Experimental results show that the federated learning model achieves high prediction accuracy, with a Mean Absolute Error (MAE) of 0.7917°C, a Root Mean Square Error (RMSE) of 1.1124°C, and a coefficient of determination (R2) of 0.9448, indicating that the model explains 94.48% of the variance in temperature prediction. These results demonstrate the feasibility of applying federated learning to greenhouse climate control, providing a balance between high prediction accuracy and data privacy. This approach has significant potential for real-world deployment in precision agriculture, providing a scalable and secure machine learning solution for distributed environments.