Use of Robust Artificial Neural Networks and ARIMA in Detecting Brief Anomalies in Gas Consumption
This paper introduces an innovative system for outlier detection that combines the strengths of an Auto-regressive Integrated Moving Average (ARIMA) model and an Artificial Neural Network (ANN). While ARIMA is traditionally used for linear predictions and ANNs for non-linear forecasting, this study demonstrates their synergistic capabilities in capturing complex, non-linear relationships between meteorological forecast variables and gas consumption patterns. The resulting system can identify anomalies, aiding building managers in reducing energy waste in HVAC systems. The process comprises two phases: first, it predicts short-term gas consumption patterns using historical data, and then it identifies outliers by detecting deviations from expected values. Remarkably, this outlier detection process doesn’t require predefined labeled examples, thanks to the system’s highly accurate gas consumption forecasts, characterized by a root mean square error (RMSE) ranging from 8 m3 to 2.5 m3.
