Data-Driven Deep-Learning Forecasting for Gas Export by FPSO: A Brazilian Pre-Salt Case

Current short-term production forecasts for oil fields deal with significant challenges due to the inherent uncertainties and volatilities of hydrocarbon production. Traditional reservoir numerical simulations often struggle to capture production dynamics at high resolution. This work proposes leveraging artificial intelligence (AI) as a tool to enhance the accuracy of gas export forecasts from Floating Production Storage and Offloading (FPSO). By exploring hidden patterns in time series data, AI enables more accurate forecasting at daily and monthly resolutions, spanning periods of up to one year. This innovative approach circumvents the limitations of traditional models, offering more reliable predictions to optimize oil and gas delivery to the market, resulting in substantial financial benefits. Data drift poses a significant challenge to the performance of neural network-based forecasting models, often preventing their deployment in production environments. Traditional AI methods, such as Gated Recurrent Units (GRU) and Autoregressive Integrated Moving Average (ARIMA), exemplify this vulnerability. This study introduces a data normalization method as a stabilizing layer for neural networks, normalizing data variance and improving distribution uniformity during model training. Two novel architectures, GRU- Norm and ARIMA-Norm, are proposed, integrating layer normalization to enhance preprocessing and align with forecasting demands. The efficacy of these models is assessed using gas export data from pre-salt platforms. Compared to baseline GRU and ARIMA models, the enhanced models demonstrated superior forecasting accuracy across the evaluated time horizons, achieving an average improvement of up to 10% in Symmetric Mean Absolute Percentage Error (SMAPE). Beyond improved accuracy, the proposed models narrow the uncertainty band, delivering more reliable predictions that closely track actual gas export trends, thereby supporting robust decisionmaking. This work offers valuable insights into mitigating data drift in production settings and provides engineers with practical, innovative tools. By advancing the state-of-the-art in forecasting, these models contribute to the literature and empower practitioners with enhanced data-driven decision-making capabilities.