Comparison of Statistical Time Series and Regression Machine Learning Models to Forecast the TerraClimate Dataset in the Bouregreg Chaouia Basin

Abstract

Water resources management relies heavily on hydrological forecasting and advancements in machine learning (ML), enabling more possibilities to improve modeling capabilities. The present research examines the effectiveness of the multiple forecasting methods applied to the TerraClimate dataset of the Bouregreg Chaouia Basin. Two different types of forecasting techniques, namely statistical time series and machine learning algorithms, are developed, and the best-performing method for the given case study is determined. The performances of naïve statistical forecasting methods and autoregressive moving average methods (ARMA) are compared with regression machine learning techniques namely linear regression, decision tree, and support vector regressor (SVR). This method is applied on the forecasted TerraClimate variables for interest features such as precipitation, actual evapo-transpiration, runoff-out, Temp-min, Temp-max, water deficit, wind speed, and water storage. The evaluation is based on global error metrics such as mean squared error (MSE), mean absolute error (MAE), and median absolute error (MedAE). The results clarify that both of the forecasting methods provide valuable results, but the strength of the regression machine learning method concludes that the most suitable option is to use it in the forecasting TerraClimate Dataset of Bouregreg Chaouia Basin, especially the decision tree model, which is the most powerful machine learning method with the lowest error metrics. Globally, the developed models are valuable for assessing the forecasting of climate features and could help climate researchers and decision-makers with a real-time forecast of climate features change using physical satellite parameters as input variables.