Forecasting Field Production Using Machine Learning Time Series

Abstract

Accurate forecasting of field production is an essential ingredient for effective planning, management, and control of construction resources. Traditional forecasting methods often rely on historical data and on the subjective experience of project managers that fail to account for the dynamic and unique nature of construction operations. As such, this thesis presents a generic framework to collect, mine and analyze ongoing field production data. It incorporates developing a time series machine learning based model to forecast production for the upcoming days. Also, to ensure its performance, the model was benchmarked with the traditional statistical based Autoregressive Integrated Moving Average (ARIMA) time series model and the Multi-layer perceptron (MLP) neural network that does not include the time factor. The framework was tested on a case study of excavation activities of a real-world infrastructure project for eight months. The developed machine learning time series model produced a satisfactory performance to forecast field production, and it showed improvements over the statistical ARIMA time series model and the MLP model with no time lag. The proposed framework can assist project managers to accurately forecast field production in order to make data-driven resource allocation decisions.