Electric utility companies in developing regions lose millions of dollars due to unplanned power outages. Particularly, in Honduras only, energy losses reach 38%, which are among the highest in Central America, with unscheduled outages accounting for nearly 30% of the total offline time. Currently, in Honduras and neighboring regions the outages are addressed in a reactive manner, where crews are dispatched once the outage occurs and that prolongs downtime, disrupts businesses, and deteriorates the quality of life, which is exacerbated due to a limited infrastructure in terms of roads and highly mountainous terrain.

Lepa Wiran AI is a relatively low-cost solution that is fed on data that is used to provide predictive information about power outages at specific grid locations, within the reach of the weather service. This system integrates freely available data from three sources: ESA's Sentinel-2 mission (10-60 m resolution), topographic data from Shuttle Radar Topography Mission (SRTM, 30 m resolution) and meteorological data from ERA5-Land reanalysis to train two classifiers. From the study, 15 environmental features were found to have the strongest predictive power, and these are the ones used to determine the risk index of a certain region. The classifiers were originally trained over a 14-day window preceding each forecast date; however, it can be retrained using different configuration lines as it best adapts to a given region.

The final model combines XGBoost and Random Forest machine learning algorithms into a soft-voting, 50/50 classifier, and it was trained on 21,902 labeled events collected across 83 substations in Honduras. The output is a decimal number between 0 and 1 for a given location, and an interactive web-based frontend visualizes the sores on a map, which also allows filtering by region and level of acceptable risk by adjusting the threshold, from a sensitive point to a more conservative index (e.g., 0.7).

The main novelty of this solution is that this system is adapted to data-scarce developing regions like Honduras, without needing stringent changes for its use. It does not require SCADA integration or proprietary sensors, and it simply relies on freely available data. Thanks to the adjustable threshold feature, it allows the companies to balance recall and precision based on seasons and resources availability.

This is a software-only solution at the moment, that can be improved through the incorporation of direct physical inspection by integrating precise sensing infrastructure or automated guided vehicles such as drones. Currently, the trained classifiers run from a local flask server that can be easily deployed from the cloud to be accessed anywhere else in the world. There is a light, responsive frontend built as a single page which also can create reports, and it accesses real time data via APIs: Google Earth Python API and weather data via Open-Meteo.

The main market are the electric utility companies. Other markets include transportation and telecommunications. Insurance companies could have some interest in the index risk.

Finally, this solution improves quality of life by reducing power disruptions due to empowering utility companies to better allocate their resources.