Improving thermal substation inspections utilising machine learning

Periodic thermal imaging inspection of air-insulated substations can lead to false negatives due to the heating effects of solar radiation and cooling effects of wind and precipitation. This work aims to characterize the effects of wind on thermal images of thermally loaded equipment, allowing thermal response forecasts to be made. Data is collected in two load patterns from an indoors experiment, comprising a current loop of two overhead- line conductors energized by a high-current DC power supply. Wind is emulated by an industrial fan. Infrared images, environmental data (ambient temperature, humidity, pressure, wind speed and direction) and electrical load data are all captured periodically. A further dataset from an in-service substation is used. Models are created using vector autoregressive and long short-term memory recurrent neural network models in order to further develop the methods presented by Bortoni et al. The results display a clear improvement over those found in the literature, highlighting the utility of modern data-processing techniques. These results present an opportunity to extract meaningful information for long term thermal condition monitoring of power substations.