Decomposition combining averaging seasonal-trend with singular spectrum analysis and a marine predator algorithm embedding Adam for time series forecasting with strong volatility

It is challenging to predict the time series data with strong volatility. Aiming to deal with this issue, we propose an innovative hybrid forecasting framework called Temporal-Frequency Reconstruction (TFR) in this study. In this framework, the averaging operation is incorporated in seasonal-trend decomposition using Loess (STL), and uniform grouped singular spectrum analysis is developed. A novel decomposition algorithm is constructed by combining these two methods, referred to as aSTL-UGSSA. Firstly, the time series data is decomposed by aSTL-UGSSA to extract latent structure information. Then, these decomposition terms are predicted by gated recurrent unit (GRU) models. To improve the prediction accuracy, a novel marine predator framework embedding Adam algorithm (MPAdam) is proposed to optimize the parameters of GRU models. Further, we analyze the factors contributing to the strong performance of TFR. TFR can not only capture the trend and seasonal signals but also effectively extract information from the remainder component. MPAdam overcomes the problem of initialization sensitivity and converges rapidly. In the short-term and long-term forecasting tasks for single-seasonal and multi-seasonal data, TFR has outperformed the state-of-the-art time series forecasting models by a significant margin.