Machine Learning Approaches for One-Day Ahead Soil Temperature Forecasting

Abstract

Present study investigates the capabilities of six distinct machine learning techniques such as ANFIS network with fuzzy c-means (ANFIS-FCM), grid partition (ANFIS-GP), subtractive clustering (ANFIS-SC), feed-forward neural network (FNN), Elman neural network (ENN), and long short-term memory (LSTM) neural network in one-day ahead soil temperature (ST) forecasting. For this aim, daily ST data gathered at three different depths of 5 cm, 50 cm, and 100 cm from the Sivas meteorological observation station in the Central Anatolia Region of Turkey was used as training and testing datasets. Forecasting values of the machine learning models were compared with actual data by assessing with respect to four statistic metrics such as the mean absolute error, root mean square error (RMSE), Nash-Sutcliffe efficiency coefficient, and correlation coefficient (R). The results showed that the ANFIS-FCM, ANFIS-GP, ANFIS-SC, ENN, FNN and LSTM models presented satisfactory performance in modeling daily ST at all depths, with RMSE values ranging 0.0637-1.3276, 0.0634-1.3809, 0.0643-1.3280, 0.0635-1.3186, 0.0635-1.3281, and 0.0983-1.3256 degrees C, and R values ranging 0.9910-0.9999, 0.9903-0.9999, 0.9910-0.9999, 0.9911-0.9999, 0.9910-0.9999 and 0.9910-0.9998 degrees C, respectively.