Application of M&V 2.0 techniques to identify and quantify non-routine events

Abstract

Energy efficiency measures (EEM) are implemented on mines to reduce energy consumption. The impact of an EEM must be determined to evaluate if any energy savings were achieved. This is done by applying measurement and verification (M&V) principles. However, unanticipated operational changes known as non-routine events (NRE), inhibit accurate quantification of an EEM. One challenge in industry is to rapidly identify an NRE. Further investigation on mines brought to light that NREs are detected and quantified through manual monitoring and strategy application. This process may take days or weeks and can lead to energy over expenditure or increased costs. Mines have implemented advanced metering infrastructure that continuously expand databases. To address the big data availability, M&V 2.0 emerged over the last decade. This field has seen little contribution in terms of anomaly detection in mining equipment. In this study, a data driven approach to rapidly identify NREs in mines is investigated. The literature study identified the importance of data preparation to expose trends and gain insight. The methodology included the organisation of mining energy data according to Eskom time of use (TOU) periods. Changepoint detection strategies were then applied to each respective TOU category. Research stipulated that the cumulative sums (CUSUM), a prescribed M&V technique, was suitable to identify anomalies in large data sets. To contribute to the field of M&V 2.0, this study compared two changepoint detection algorithms to the CUSUM method —pruned exact linear time (PELT) and binary segmentation (BinSeg). A package called changepoint, developed in the R programming language, was applied to implement the BinSeg and PELT algorithms. It was found that the BinSeg and PELT algorithms identified more changepoints than the CUSUM method. Overall, PELT was the most accurate. One flaw in the CUSUM method was that it did not identify changepoints in data with high variability. Whereas, the PELT and BinSeg algorithms may have been inaccurate if the parameters were not correctly specified for each data set. One such parameter was the minimum segment length, which stipulated the minimum duration of a valid changepoint. In this study, the minimum segment length was generalised to mining energy data as a whole and should have been specific to each data set. In conclusion, the algorithm driven technique was proved to be effective in identifying NREs in mining energy data. Such algorithms can be applied to every data stream in the mine. The impact of this is more rapid identification of equipment failure or increased energy consumption. In this study, energy data from the national lockdown was used. During this period, increased demand in the gold plant energy data was identified. In a time where the mine could not generate revenue, the increased energy on the gold plant was estimated to cost R27000 daily.