Multistage decomposition transformer network for predicting complex long time series of heavy oil parameters

Abstract

The primary indicators of heavy oil production include temperature, pressure, and various other factors, which exhibit rapid trend changes, erratic wave patterns, and irregular long-term behaviors, severely hindering accurate predictions. To effectively capture the long-term nature and complexity of heavy oil parameters, we propose a multistage decomposing transformer network (MDTN). The MDTN consists of two non-autoregressive decoders, an encoding structure with time encoding, and a time series parser. In this paper, we introduce a time series decomposition (TSD) strategy that breaks down complex long-time series into two simpler trend components and residual components. For the long-term analysis, we employ a local sensitive hash attention mechanism to further decompose these two components into multiple subsequences, followed by self-attention calculations for each subsequence. Additionally, to enable the model to fully leverage the temporal information of the sequence, we embed time, value, and position into each input layer of the encoder. To achieve rapid predictions and minimize error accumulation, we have designed a novel non-autoregressive decoder. Finally, the two sequences are combined through a convolution layer. A substantial number of experiments conducted on heavy oil parameter datasets and publicly available datasets demonstrate that the proposed method yields optimal results. For instance, in the complex long-term prediction of boiler temperature, the MAE value of the proposed method reaches 0.715 at the 1008 prediction step, which is nearly 0.1 lower than that of alternative methods.