A Granularity Invariant Method for the Classification of Energy Production Time Series in Photovoltaic Plants

The treatment of photovoltaic power production time series often faces the challenge of unifying the granularity of the series when generating a predictive model. This can limit the generation of a dataset in terms of the time covered and the number of examples. In addition, models built with data of static granularities tend to show rigidity when facing granularity variations, invalidating them for scenarios different from those of the data they were trained on. To address this issue, this paper presents a novel method specifically indicated for Deep-Learning models that shows invariance to granularity called Synthesis. This operation can be added as a layer to an artificial neural network, allowing it to be applied to any power production time series and synthesizing the content of an arbitrarily long time series into a fixed-size vector which can be used for classification or regression regardless of the initial time series length. The experiments with the NIST Campus Photovoltaic dataset demonstrate the effectiveness of the method, showing an F1-Score of 1.0 for the classification of series with granularities between 2 minutes and 2 hours, and an F-Score of 1.0 for the classification of time series with variations of granularity throughout time when training with 5-minute granularity samples.