A methodology to determine the optimal train-set size for autoencoders applied to energy systems

In the latest years, deep learning has been massively used to face problems that have not been solved by means of classical approaches. In particular, an autoencoder is a popular unsupervised artificial neural network that learns efficient data representations (encoding) by training the network to ignore features with a small content of information. Even though autoencoders over-perform classical techniques in several applications like anomaly detection, dimensionality reduction, features denoising, and missing values imputation, the literature does not provide a commonly accepted methodology to define the optimal amount of data needed to train the model. This paper proposes a procedure to determine the optimal train-set size to minimize the reconstruction error of an autoencoder with pre-defined structure and hyper-parameters that will be trained to encode the normal behavior of energy generation systems. This procedure exploits the outcome of learning curves, a powerful tool to track algorithms performance while the train-set dimension varies. Afterward, the procedure is applied to three real case studies where two types of autoencoders are trained to learn the normal behavior of a YANMAR combined heat and power unit with the scope of detecting incoming anomalies. In the end, the outcomes of the procedure are explained and, under the constraint of a daily retraining frequency, 6 weeks are identified as the optimal train-set size for both autoencoders.