Semantically Enriched Multi-level Sequential Pattern Mining for Exploring Heterogeneous Event Log Data

Photovoltaic (PV) event log data are typically underexploited mainly because of the heterogeneity of the events. To unlock these data, we propose an explorative methodology that overcomes two main constraints: (1) the rampant variability in event labelling, and (2) the unavailability of a clear methodology to traverse the amount of generated event sequences. With respect to the latter constraint, we propose to integrate heterogeneous event logs from PV plants with a semantic model of the events. However, since different manufacturers report events at different levels of granularity and since the finest granularity may sometimes not be the right level of detail for exploitable insights, we propose to explore PV event logs with Multi-level Sequential Pattern Mining. On the basis of patterns that are retrieved across taxonomic levels, several event-related processes can be optimized, e.g. by predicting PV inverter failures. The methodology is validated on real-life data from two PV plants.