A Comparison of Machine Learning Algorithms for Thermal Rating Calculations of Power Cable Systems Based on Measurement Data

Abstract:
The established analytical calculation method for the steady state thermal rating of power cables is given by the IEC 60287. Anyhow, due to a general increase in complexity in energy systems and the speed of the transformation process, a demand-pull arises in the energy sector for new sophisticated digital solutions, like the application of machine learning. Therefore, 3 different machine learning algorithms were used to derive mathematical models based on measurement data, for the prediction of the cable surface temperature. The used data was recorded over a time period of two years, in which multiple steady state current scenarios were applied on a 400-kV-cable test setup with a real urban laying profile. To not only compare the results of the mathematical models, which were achieved with the different machine learning algorithms, with each other, a reference steady state thermal rating calculation according to IEC 60287 was made in addition. It could be shown that accurate mathematical models with accuracies in the range of 97 to above 99% can be derived already on the basis of relatively small data sets. Furthermore, it could be shown on an example, how a neural network can be used to predict the cable surface temperature for an entire consequent year.