Multivariate Production Forecasting Via Deep-Learning Approach

Summary Factual production assessment is the bottleneck for field integrity management that constraints the efficient development of oil and gas fields. Forecasting gas production is a challenging task because of the heterogeneity of sediment, the behavior of complex fluids, and tortuous flow channels. However, traditional forecasting approaches such as decline curve analysis and numerical simulation have the potential to represent the time series data. Still, they are not capable of handling complex features of the reservoir. Due to these limitations of traditional approaches, data-driven approaches (machine learning or deep learning) are used nowadays. In this context, a novel forecasting model for gas production is proposed in this work. This study used the gas production data from Equinor's Volvo Field. After collecting datasets, the first step is to clean the data and remove the unwanted data, missing values, rows, and columns. Then, split the dataset into a train set (75%) and a test set (25%). The proposed system offers a logical solution to the challenge of forecasting multiple data points, and its sliding window mechanism ensures that no data leakage occurs during training. We have also developed deep-learning forecasting models (gated recurrent units (GRU), long short-term memory (LSTM), and temporal convolutional networks (TCN)), stacking the layers to concentrate on various timeframes and evaluating them against various commercially available approaches. The TCN model indicated the best performance compared to other deep learning methods. This approach is trained on different Volvo field datasets to assist in gas production.