Flowvision AI Context Aware Production Forecasting Via Virtual Flow Metering and Operational Text Mining

Abstract

Abstract This study, completed during the 2025 GOTECH® hackathon, presents a data-driven framework to identify production-optimization opportunities and to forecast incremental production. The approach was applied to an offshore platform with wells operating under high, low, and very low pressure, where planned and reactive activities are executed to manage challenges that can cause downtime, deferred production, and inefficiencies. Key issues included wax deposition in tubulars and gas-lift inefficiencies. The study targets (i) identification of an effective wax inhibitor to reduce slickline interventions, (ii) optimization of gas-lift injection per well, and (iii) estimation of incremental production resulting from these activities via advanced forecasting. Methods, Procedures, Process Historical SCADA/PI datasets and maintenance/operations records were curated across the platform. Phase-specific virtual flow metering (VFM) models for oil, gas, and water were built using tree-based regressors, enabling the estimation of instantaneous production rates in periods with limited or sparse well‑test data. In parallel, large language models (LLMs) were employed to perform domain‑adapted text classification on daily operational reports and intervention tickets. This workflow enabled automated identification and time‑stamping for activities such as pigging, slickline‑based wax removal, chemical‑inhibitor dosing, gas‑lift set‑point adjustments, choke‑setting changes, and operational‑flexibility events. These contextual interventions were subsequently engineered into covariates—including event flags, decay functions, and dwell‑time metrics—to enrich the predictive feature space. A transformer‑based forecasting architecture then generated rolling five‑day production forecasts using a fifteen‑day lookback window. The model consumed both the VFM‑derived flow‑rate estimates and the engineered contextual features, providing a unified, data‑driven framework for short‑term production prediction under dynamic operational conditions. Results, Observations, Conclusions Model performance was evaluated using Root Mean Squared Error (RMSE) and Mean Average Error (MAE) for thermodynamic state variables (tubing-head pressure, casing-head pressure, choke temperature) and for phase rates (oil, gas, water) on train, validation, and test splits. Forecasting skill was assessed under a rolling five-day horizon with a fifteen-day lookback by comparing predicted values against observed production. Additional evaluation included intervention-classification accuracy from LLM outputs and the backtesting of incremental-production estimates against historical operational changes. Results showed that the integrated pipeline demonstrated higher accuracy on state variables than on multiphase rates; oil forecasts outperformed gas due to variability induced by gas-lift operations. Incorporating contextual features from LLM-extracted interventions stabilized predictions and reduced error variance relative to sensor-only baselines. Backtesting demonstrated improved short-horizon reliability during disruptions such as pigging and wax-removal campaigns, confirming the value of converting unstructured metadata into predictive covariates to enhance production forecast accuracy. Novel/Additive Information The proposed contribution is an end-to-end, deployment-oriented pipeline that integrates unstructured operational text with VFM-derived signals to enable production forecasting. Automated event mining turns "unused metadata" in reports into actionable covariates, enhancing short-horizon skill where physics-only or sensor-only methods struggle. A strong emphasis on back testing, interpretable context features, and lightweight visualization modules support scalable adoption in daily production meetings and exception-based surveillance workflows. This integrated design enables operational teams to leverage both engineered signals and domain‑specific narrative insights within a unified, data‑driven forecasting framework. AI Disclosure: AI and machine-learning tools were used to support this study. Tree-based regression models were used to develop VFM estimates of phase rates; a large language model was used to classify and time-stamp interventions from unstructured operational text; and a transformer-based time-series forecasting model was used to generate rolling production forecasts and to support incremental-production backtesting.