A Comprehensive Time Series Forecasting for Motor Winding Temperatures

Predictive maintenance of electric motors in large refineries is a challenge. Electric motors supply the power to key process equipment such as pumps and compressors, and their failure can lead to large financial losses. As such, it is essential to predict future condition of these motors in advance. The failure of the electric motors has been linked to some key parameters such as weather-related factors (i.e., ambient temperature, precipitation level, etc.), a motor’s bearing and winding temperatures, as well as motor’s operating current, for which telemetry data can be collected on a minute level.This work studies and applies use of classic time series techniques such as statistical models as well as machine learning solutions, post intensive data engineering and exploratory data analytics. Several different algorithms were studied and deployed to address both data engineering and data science scope of the work in an automated fashion to be implemented in machine learning operationalization. Eventually, a hybrid model (combined statistics and machine learning) was developed, tested and deployed on several electric motors with versatile attributes, and the models showed very good precision on the hold-out test data, with a long predictive power.Selected models are used to generate long term (up to six months) forecast for the average daily highest winding temperature.