Decline Curve Analysis with LSTM: A Practical Guide

Decline curve analysis (DCA) is a cornerstone of reserve estimation and production forecasting in petroleum engineering. Traditional hyperbolic and exponential decline models assume smooth, predictable decline trajectories. In reality, production data is noisy, interrupted by well interventions, and often shows nonlinear behavior that classical models struggle to capture.

Long Short-Term Memory (LSTM) networks are a type of recurrent neural network specifically designed to learn patterns in sequential time-series data. Unlike classical statistical models, LSTMs can adapt to changes in decline behavior and discover hidden relationships in the data.

We trained our LSTM model on historical production data from over 500 producing wells. The network learns to distinguish between normal decline, temporary production loss due to maintenance, and actual changes in well productivity. It can forecast 12 months ahead with a mean absolute percentage error (MAPE) of just 8.3%—compared to 15.2% for traditional hyperbolic decline models.

This guide covers the fundamentals: how LSTM cells work, how to prepare production data for the network, and how to interpret the confidence bands in our forecasts. We'll walk through a real example from an unconventional play, showing how the LSTM-based forecast revealed a sub-surface recompletion opportunity that traditional DCA would have missed.

Using our Production Optimization legacy app, geoscientists and engineers can move beyond spreadsheet-based reserve calculations and leverage machine learning to extract value from their historical production databases.