Synthetic Images Generation for Deep Learning Assessment Towards the Inflow Forecast for Power Systems Operation

Thaís Rocha¹; Ieda G. Hidalgo¹; André Franceschi de Angelis¹

¹ Unicamp - University of Campinas

Resumo

This paper presents the reasons and the means to work with synthetic images sequences to assess Deep Learning (DL) Artificial Neural Networks (ANN), aiming a better technology for water inflow forecast to base the operations of wide power systems. We developed a computer program that efficiently generates the images and the respective numeric data series, allowing the training of ANNs. As the program includes an user-defined displacement between the image and the numeric values when it creates the sequences, the prediction ability of DL ANNs can be evaluated over temporal data series in several fields, even in noisy scenarios. We ended up with a powerful software tool, extensible and portable, able to incorporate more image generators algorithms. Our main contributions are exploring the DL potential to make predictions in temporal image data series and developing a tool for this purpose.

Palavras-chave: Hydroelectric plants; Temporal data series; Climate forecast; Image analyses; Data assimilation