Data-driven approach for temperature and relative humidity forecasting in solar dryer dome facility

Abstract

Accurate prediction of temperature and relative humidity is essential in Solar Dryer Dome (SDD) facilities to ensure strict quality of the agricultural products. Compared to traditional control models, ML-based approaches provide benefits in terms of adaptability, as they are capable of capturing complex and non-linear patterns in time series data. In this project, primary time series datasets collected from SDD facility in North Jakarta were used to forecast the next hour temperature and relative humidity values within the 5-minutes resolution. There were three sensors placed in three different locations: outside the facility near the front door, inside the facility near the front door, and near the fan inlet. The collected data showed daily seasonality patterns with temperature and relative humidity possess an inversely correlated relationship due to their inherent physical interaction. On the modeling part, this research compared three different predictive models such as XGBoost, Linear Regression, and Facebook’s Prophet based on the RMSE and R² scores on the test set. XGBoost and Linear Regression provided on-par performance, with XGBoost showing its best at temperature prediction (RMSE: 1.47°C, R²: 0.96) and Linear Regression showing its best at relative humidity prediction (RMSE: 3.20%, R²: 0.96). Meanwhile, Prophet exhibits the lowest performance in both variables with RMSE: 3.56°C and R²: 0.72 for temperature, and RMSE: 8.42% and R²: 0.76 for relative humidity forecasting. Despite the low performance, the Prophet model was advantageous in terms of consistency across horizons, while the two best models show decreasing performance in higher horizon settings. These findings suggest the promising application of ML techniques in SDD facilities to support sustainability in agro-industrial business.