Detecting potential wasted interventions on customer premises with natural language processing

By managing 95% of the power distribution network in continental France, Enedis operates interventions for 32 million residential customers on behalf of their electricity suppliers. Although more than 90% of these customers are equipped with Linky meters, several million on-site interventions are still necessary every year to meet their needs. Some trips to intervene on Enedis customer premises at the request of their electricity supplier lead to wasted interventions. Besides subsequential customer dissatisfaction, field operators and programmers at Enedis are needlessly mobilized, not to mention the trip’s CO2 impact. This article describes how we are working on avoiding some of these wasted interventions. The system developed by Enedis includes an innovative approach based on the analysis of textual comments provided by the supplier with their intervention request, especially when the content suggests a discrepancy between the type of intervention requested and the need expressed by the customer. These interventions are therefore suspected of being misclassified, and field operators may not be able to carry them out successfully. This analysis uses natural language processing models based on an algorithm for French language processing enriched with Enedis' technical vocabulary. Every morning, the system provides employees in charge of managing activities with a list of these interventions likely to be wasted, so they can call the customer back, reclassify the interventions or cancel them. In 2023, more than 230 users have used our system, thus avoiding 65,000 wasted or unnecessary interventions on a national scale, corresponding to a €3.7M gain (costs of the unnecessary trips to client premises).