An Agentic Framework for Causal Discovery and Forecasting in Oil and Gas Time Series

Ahmed Esmin, Alexandre Mello Ferreira, Anderson Rocha

Presentation Attendance: Yes, we will present in-person

Keywords: Agentic frameworks, Causal discovery, Time-series forecasting, Foundation models, Decision support

TL;DR: An agentic framework that integrates causal discovery and foundation-model forecasting to support operational decision-making in oil and gas time-series analysis.

Abstract:

Industrial time-series analysis requires both accurate forecasting and actionable
explanations, particularly in production systems where interventions propagate
with delays across interconnected assets. We present an ongoing applied project
that develops an agentic framework integrating time-series causal discovery, tem-
poral dependency analysis, and foundation-model forecasting into a unified work-
flow for operational diagnosis and decision support. Engineers interact through a
conversational interface to (i) localize operational events and regime changes, (ii)
estimate causal links and time lags among injector and producer variables, and
(iii) generate short- and medium-horizon forecasts using a time-series foundation
model (TimesFM). We demonstrate the approach on realistic simulated oil and
gas production datasets generated with Unisim, capturing injector–producer inter-
ference patterns. Early observations indicate that combining causal structure with
foundation-model forecasting improves interpretability and supports faster inves-
tigation than forecasting-only baselines, while maintaining competitive predictive
performance. We summarize practical lessons from integrating causal tools and
forecasting into an engineer-facing workflow and outline next steps toward multi-
scenario causal discovery and interactive deployment.

Track: Industry and Applications Track (max 2 pages)

Submission Number: 29