Transformer Load Losses prediction by means of Finite Element Method Simulations and Machine Learning

Transformer end-users but especially for all connected sustainability impacts they certainly have. Load losses are measured during the well-known factory acceptance test (FAT) called load losses test ([1], [2] and [3]). Load losses are the energy loss that, together with no-load losses, mostly impact power transformer efficiency. A good prediction of load losses avoids oversized units or rejection of the units for not passing the FAT test. Legacy tools including algorithms based on finite element method simulations (FEM) ([4]) and regressions trained considering the data available in the organization are already used. This paper presents the last trial to improve the results of these already available legacy algorithms. A dataset of around 350 unique power transformer designs (more than 600 tested transformers considering sister units), from 13 different factories inside the organization, has been used so far but the dataset is in continues enlargement because of the successful results. The main news of the algorithm includes a massive usage of detailed 3D FEM simulations to train the tool and the test of multiple machine learning (ML) techniques. 3D FEM simulations are used only for the training of the algorithm and must not be performed to predict load losses in new units expect for some special-purpose transformers (e.g. industrial transformers) as defined in [4].