Temporal Pattern Recognition and Long-Term Forecasting of Gas Flaring in Nigeria: A Comparative Evaluation of LSTM, Random Forest, and LightGBM

Abstract Nigeria's gas flaring problem is neither simple nor new. Despite decades of regulations and technological advances, flaring persists across the Niger Delta, bringing major economic losses and environmental harm. With growing pressure from domestic climate pledges and international partners, the ability to forecast flaring trends and assess policy impacts has become critical for infrastructure planning and national commitments. This study uses a fifteen‑year monthly dataset (2002–2017) from NNPC Annual Statistical Bulletin and NUPRC Data Repository, 331 observations across seven features. Preprocessing confirmed that produced gas and fuel gas are the strongest drivers of flaring. Three machine learning models were developed: Long Short‑Term Memory networks, Random Forest, and Light Gradient Boosting Machine. LSTM used a two‑layer, 96‑unit design with a three‑step input window. Random Forest and LightGBM were tuned via Bayesian methods. Data was split 70/15/15 for training, testing, and validation. Model performance was assessed using the coefficient of determination (R2), mean absolute error (MAE) and root mean squared error (RMSE). All three models performed well on the validation data. LSTM returned an R2 of 0.52 with an MAE of 11.29, and an RMSE of 15.55, confirming its ability to extract meaningful temporal patterns despite being the worst in performance, as the outlier capping was not favorable to the model. Random Forest achieved an R2 of 0.90 with an MAE of 3.23, and an RMSE of 7.0, demonstrating that it can handle the nonlinearities embedded in the features. LightGBM delivered an R2 of 0.91, with an MAE of 3.86, and an RMSE of 6.90, striking a useful balance between accuracy and computational cost. Beyond predictions, the analysis revealed cyclical flaring patterns and nonlinear regime shifts tied to major policy interventions. Scenario projections to 2030 indicate that while current paths fall short of zero‑routine‑flaring targets, redirecting export‑oriented gas to domestic industrial use could speed reductions significantly. This work provides a reproducible, multi‑model framework for decoding flaring dynamics, offering policymakers a data‑driven tool for infrastructure and timing decisions, operator's clearer planning insights, and the broader energy transition effort a demonstration that forward‑looking foresight is within reach.