Use of Several Sources of Spatio-temporal Information to Improve Short-term Photovoltaic Power Forecasting.

In recent years, the share of photovoltaic (PV) power in Europe has grown: the installed capacity increased from around 10 GW in 2008 to nearly 185 GW in 2021. Due to the intermittent nature of PV generation, new challenges arise regarding economic profitability and the safe operation of the power network. To overcome these issues, a special effort is made to develop efficient PV generation forecasting tools.

Several sources of information are currently investigated in the literature. Each one possesses different characteristics, which make them horizon-specific. For short-term forecasting (i.e. from a few minutes to 6-hour ahead), endogenous inputs, namely past PV production measurements, are typically the main drivers. With the development of PV plants, and the advances in smart monitoring and measurements, we observe a paradigm shift from temporal- to spatio-temporal (ST)-based forecasting models. This family of models considers features that exploit ST correlations in the data, such as observations from spatially distributed portfolios of PV plants. This new paradigm offers power producers the possibility to economically value information from geographically distributed plant networks in the form of forecast accuracy improvements, and prepares the ground for a data-sharing market.

Depending on its distribution or density, a PV network may partially account for the complex ST processes at stake (e.g. mainly sites located upwind or crosswind). To fill this gap, satellite-based observations are an appealing option. With recent developments, geostationary satellites can capture images of Earth at a temporal resolution of less than an hour, which enables operational uses. Contrary to the spatial inflexibility inherent to PV networks, satellite-based observations offer the possibility of covering the whole vicinity of the site location, and much more. In that context, relevant features selection tools need to be considered.

In this work, we propose the following contributions to the state of the art. Traditionally in the literature, observations from spatially distributed units and satellite-derived information are used separately. We propose an incremental approach to assess the impact of one or several sources of data on the forecasting performances of the considered regression model. This approach shows that the combination of various sources of ST information leads to higher accuracy than when inputs are considered individually. This is assumed to result from the difference in spatial resolutions of both features. In this specific case study, we highlight the limits of the PV plants portfolio through an analysis of the local topography and wind distribution at several altitudes Then, we consider cloud opacity maps obtained from infrared channels. Despite being under-represented in the literature (only two studies have been found), infrared channel-based data present the advantage of offering nighttime observations of cloud cover, which contributes to improving early morning forecasts.

The proposed approaches are evaluated using 9 PV plants in France and for a testing period of 12 months.