Imputation of missing measurements in PV production data within constrained environments

Abstract

The inability to measure energy production accurately in photovoltaic power stations is an important issue affected by photovoltaic plants. In certain contexts, the imputation of missing values can be complicated given the lack of similar measurements from other environments and/or the difficulty of generating predictive models when there are simply too many unknowns, when other measurements are not available, or when there are no external correlated signals (e.g., solar irradiation, ambient temperature, modules temperature or wind speed). For such limited environments, this paper proposes a solution based on an artificial neural network with an architecture built in the form of an encoder–decoder. The proposed model can be trained with time series of energy production noting missing data of any kind. This paper demonstrates that the proposed model can perform imputations of missing data comparable to those made by auto-regressive and clustering techniques, even when the time series to be predicted is the same as the one used for training. The model is well suited for environments where the severity of missing values range between 30% and 70%, being especially indicated for environments where the severity of missing values reaches 50%, which surpasses in all the evaluated metrics other methods.