Attention-based LSTM network-assisted time series forecasting models for petroleum production

Petroleum production forecasting is the process of predicting fluid production from the wells using historical data. In contrast to the traditional methods of analysing surface and subsurface parameters governing fluid production, machine learning (ML) techniques are being applied to forecast the production. The major drawback of traditional and conventional ML techniques is that they are time-consuming and often lack good forecasting power. In this work, time-series forecast models based on powerful and efficient ML techniques are developed to forecast production with historical data. We have fused the attention mechanism into the long short-term memory network, which is referred as the attention-based long short-term memory (A-LSTM) network. The A-LSTM network is fast and accurate, thus solving the low forecasting power problem. To ensure no data leakage occurs during training, and to build a reliable data-driven forecasting approach, we construct the dynamic floating window with varying window sizes over the entire production data. The dynamic floating window slides one-step forward after every prediction and continues till the last production window enabling the model to fit the new data automatically. We have tested and validated the proposed forecasting models with the ML algorithm using actual production data for three wells from entirely different geographies. We then compared them with statistical, deep learning, hybrid, and ML approaches. The genetic algorithm (GA) is applied to optimize the hyper-parameters of the A-LSTM. The results of a comparative analysis show that the A-LSTM network statistically and computationally outperforms the other models for forecasting petroleum production. [Display omitted] •Attention-based Long Short-Term Memory (A-LSTM) was used for forecasting the production.•A genetic algorithm (GA) was used to optimize the hyper-parameters of A-LSTM.•Seven forecasting models, consisting of several inputs and output were developed and tested.•A well-trained A-LSTM algorithm was observed superior, easy, and computationally efficient.