An introduction of a reward-based time-series forecasting model and its application in predicting the dynamic and complicated behavior of the Earth rotation (Delta-T values)

Abstract

This study aimed to propose a reward-based time-series forecasting model (RBTM) as a novel method to model and forecast time series with no learning algorithm. This method extracts features of time series as distinctive rules and calculates their proper relevant rewards. These rules and rewards are kept in the knowledge base part. The next value is estimated by using rules, rewards, and the prior value of a time series. To evaluate RBTM, it was run to forecast four different time series, and the results were compared with those of MLP, TDNN, NARX, ANFIS, and LSTM models. The findings showed that errors in RBTM decreased by 0.6% to 84.4% compared to the other methods. Then, RBTM was used to forecast the dynamic and complicated behavior of the Earth rotation (DT values), as a real-world application. In this case, the data from 1800–2000​ and 2000–2018 were selected to model and test the parts, respectively. The mean absolute error (MAE) and root mean square error (RMSE) of the test part were 1.04 and 1.40, respectively. To evaluate the accuracy of the forecasts, the RBTM results were compared with the prediction results of the previous models. The findings indicated that the MAE and RMSE of RBTM decreased by 48% and 39%, respectively, as compared to those of the other methods. Finally, the results of forecasting DT values using RBTM from 2019–2030 were reported.