Multivariate Models for Photovoltaic Power Forecasting with Non-climatic Exogenous Variables

Forecasting electricity generation from renewable resources is crucial for the efficient planning and operation of power systems. The development of forecasting models based on local meteorological variables is common, however, sometimes this information is unavailable. This study explores the use of multivariate models that do not incorporate meteorological variables, but use historical power-generated data from eight PV plants located in the same region to predict the future value of a target plant. This allows for improved forecasting when meteorological variables are unavailable and the only information available is the generation of the PV plants. The performance of LSTM and BiLSTM networks is compared for different time horizons, considering various lags of the power series itself for estimating future values. The main contributions of this study include the introduction of power time series from other plants as model inputs, the use of spatial interpolation to fill in missing data and the application of causality tests between time series for the selection of predictor variables, and the uncertainty associated with the predictions is analyzed using quantile regression techniques.