Open Collections

Abstract

Regulatory agencies depend on accurate forecasting of oil and gas well production to support planning, decision-making, and risk assessment. Traditional Decline Curve Analysis (DCA) models are commonly used due to their simplicity, but they rely heavily on long production histories. As a result, they often fail to produce reliable forecasts for new wells with limited data. These models are also manually fitted, introducing human bias and inefficiency. This thesis proposes a hybrid DCA model that combines machine learning (ML) models with DCA to improve accuracy, especially during early production of wells. We first introduce Auto-DCA, an automated method that selects the optimal start date for DCA fitting and then fits the DCA model with minimal error, removing the need for expert input. We then train the models: Random Forest (RF), Gradient Boosting Regressor (GBR), and Neural Hierarchical Time-Series (NHITS), on clusters from older wells with long production histories. These models are used to generate synthetic data extensions for new wells. Auto-DCA model is then fitted on extended data to create hybrid DCA models: RF-DCA, GBR-DCA, and NHITS-DCA, which are then evaluated against Auto-DCA. The thesis also explores how different clustering strategies affect model performance. We test three methods: formation-based clustering, K-means clustering, and k-Nearest Neighbour (kNN) based clustering. The results show that all hybrid DCA models outperform the Auto-DCA model, particularly with only one year of input data. RF-DCA and GBR-DCA perform consistently across cluster types, while NHITS-DCA performs best when trained on clusters rich in formation-specific data.