Developing machine learning models for predicting daily relative humidity and solar radiation using lagged time series data inputs in a semi-arid climate

Abstract

Estimating relative humidity and solar radiation is crucial for understanding their impact on the hydrological cycle, which in turn affects water resource availability and distribution. Accurately predicting these variables is challenging due to their non-linear behaviour. Machine learning (ML) techniques have attracted significant attention for addressing such complex non-linear problems. In this study, the prediction of relative humidity and solar radiation for ICAR-IARI (Indian Council of Agricultural Research - Indian Agricultural Research Institute), New Delhi, India, under semi-arid climate, was performed using long-term data spanning 31 years (1990–2020) and developed machine learning models such as linear regression (LR), multilayer perceptron (MLP), sequential minimal optimization-support vector machine (SMO-SVM), additive regression (AR), and random forest (RF). The performance of these models was evaluated using various statistical metrics, including coefficient of determination (R²), Nash–Sutcliffe efficiency (NSE), index of agreement (d), mean absolute error (MAE), root mean squared error (RMSE), relative absolute error (RAE), relative root squared error (RRSE) and mean absolute percentage error (MAPE). The results indicated that for relative humidity prediction, the RF model yielded the best results during training and testing periodswith statistical metrics R², NSE, d, MAE, RMSE, RAE, RRSE and MAPE as 0.79, 0.40, 0.74, 6.31, 12.66, 54.62, 74.70 and 9.90, respectively during testing phase. The SMO-SVM model emerged as the best performer for solar radiation prediction, with performance metrics during the testing phase as follows: R² = 0.89, MAE = 2.79, RMSE = 3.55, RAE = 52.54, RRSE = 59.59, NSE = 0.64, d = 0.88, and MAPE = 20.50. The findings of this study could be useful for developing and comparing relative humidity prediction models under different climatic conditions, using similar long-term data.