Time series production forecasting of natural gas based on transformer neural networks

Abstract

Time series forecasting of gas production plays a crucial role in enhancing the stability of production, optimizing development strategies, and effectively increase the life cycle of gas wells. However, the precision of these forecasts is often compromised by two primary factors: (1) the complexity and randomness inherent in production time series data and (2) the limited ability to model dependencies within temporal sequences, especially in the context of long-term, multi-step forecasts, which can lead to instability in the prediction model's results. To address these challenges, this paper introduces a novel method. Initially, Multilevel Discrete Wavelet Decomposition (MDWD) is employed to mitigate the raw gas production series' instability, complexity, and randomness. This is achieved by decomposing the input signals into their respective periodic and trend components. Subsequently, gas production modeling is executed using transformer neural networks equipped with a multi-head attention mechanism to learn sequential dependencies effectively, irrespective of the temporal distance. The architecture of this model is built upon an encoder-decoder framework. The encoder is designed to generate representations of historical gas production sequences of any length, while the decoder can generate arbitrarily long future gas production sequences. The interconnection between the encoder and decoder through the multi-head attention mechanism is a crucial aspect of this model. In two distinct experiments focusing on gas filed production data, the RMSE for one-step forecasting results produced by the proposed method was remarkably low, at 0.1911 and 0.3816, respectively. Moreover, the RMSE for 7-day multi-step predictions stood at 1.7358 and 1.2146, respectively, showcasing significant improvements over other methods. With accurate results of multi-step forecasting, this work contributes to the effective utilization of conventional and unconventional energy resources.