Prediction of length-of-day using extreme learning machine

Abstract Traditional artificial neural networks (ANN) such as back-propagation neural networks (BPNN) provide good predictions of length-of-day (LOD). However, the determination of network topology is difficult and time consuming. Therefore, we propose a new type of neural network, extreme learning machine (ELM), to improve the efficiency of LOD predictions. Earth orientation parameters (EOP) C04 time-series provides daily values from International Earth Rotation and Reference Systems Service (IERS), which serves as our database. First, the known predictable effects that can be described by functional models—such as the effects of solid earth, ocean tides, or seasonal atmospheric variations—are removed a priori from the C04 time-series. Only the residuals after the subtraction of a priori model from the observed LOD data (i.e., the irregular and quasi-periodic variations) are employed for training and predictions. The predicted LOD is the sum of a prior extrapolation model and the ELM predictions of the residuals. Different input patterns are discussed and compared to optimize the network solution. The prediction results are analyzed and compared with those obtained by other machine learning-based prediction methods, including BPNN, generalization regression neural networks (GRNN), and adaptive network-based fuzzy inference systems (ANFIS). It is shown that while achieving similar prediction accuracy, the developed method uses much less training time than other methods. Furthermore, to conduct a direct comparison with the existing prediction techniques, the mean-absolute-error (MAE) from the proposed method is compared with that from the EOP prediction comparison campaign (EOP PCC). The results indicate that the accuracy of the proposed method is comparable with that of the former techniques. The implementation of the proposed method is simple.