Due to environmental concerns, renewable energy sources (RES) play an increasingly important role in the energy mix. In France, from 2018 to 2019, an increase of 21.2% and 7.8% of energy production was observed for wind and solar respectively [1]. RES are characterized by high variability and limited predictability, mostly due to their dependence on meteorological factors. This variability presents challenges for RES integration into grids and electricity markets: as the penetration of RES increases, power system balancing becomes more complex, and congestions may occur in the grid. This lack of predictability can also have financial consequences. In some European countries, energy producers have to pay penalties proportional to the forecasting error of the injected power. To address these challenges, it is important to accurately predict the future amount of energy production. In this paper we propose a spot statistical forecasting model for very short-term time horizons (from a few minutes up to 6 hours ahead). This model is based on a combination of heterogeneous inputs with a conditioned learning approach. Spatio-temporal inputs (measurements from geographically distributed PV sites and satellite images) are used to enhance short-term predictability, while a weather analog approach enables adaptability to changes in meteorological conditions by considering the most relevant past observations. The performance evaluations are carried out on a case study featuring nine PV plants located in France, over a one-year period. Index Terms-Short-term solar power forecasting, analog approach, conditional forecast, spatio-temporal, auto-regressive processes, smart grid.

Variable renewable energy has a growing impact on electricity markets and power systems in many regions of the world. In this context, the 17th International Conference on the European Energy Market EEM20 set up a competition to develop probabilistic forecasting tools of wind production at a regional level. This paper proposes an adaptive approach for regional wind power forecasting. A physics-oriented preprocessing of the data delivers analog weather patterns and windpower-related variables, then a k-means clustering of wind farms further reduces the dimension of the problem. The generated representative features feed a Quantile Regression Forests model that produces sharp and reliable predictions. As a result, our model won the competition with a relative improvement of the average pinball loss of 6.7% and 14.7%, compared to the teams ranked second and third respectively.

Renewable Energies (RES) penetration is progressing rapidly: in France, the installed capacity of photovoltaic (PV) power rose from 26MW in 2007 to 8GW in 2017 [1]. Power generated by PV plants being highly dependent on variable weather conditions, this ever-growing pace is raising issues regarding grid stability and revenue optimization. To overcome these obstacles, PV forecasting became an area of intense research. In this paper, we propose a low complexity forecasting model able to operate with multiple heterogenous sources of data (power measurements, satellite images and Numerical Weather Predictions (NWP)). Being non-parametric, this model can be extended to include inputs. The main strength of the proposed model lies in its ability to automatically select the optimal sources of data according to the desired forecast horizon (from 15min to 6h ahead) thanks to a feature selection procedure. To take advantage of the growing number of PV plants, a Spatio-Temporal (ST) approach is implemented. This approach considers the dependencies between spatially distributed plants. Each source has been studied incrementally so as to quantify their impact on forecast performances. This plurality of sources enhances the forecasting performances up to 40% in terms of RMSE compared to a reference model. The evaluation process is carried out on nine PV plants from the Compagnie Nationale du Rhône (CNR).

In a context of natural resources depletion, weather-dependent renewable energy sources play an increasingly important role in the energy mix. Yet, high shares of renewables can jeopardise the safe operation of power grid due to their variable nature. To address this challenge, it is essential to know the future amount of energy produced to balance production and consumption. This paper aims at investigating photovoltaic generation short-term forecasting and particularly spatio-temporal approaches. These approaches permit to exploit the spatial dependency of weather variables to provide valuable information regarding cloud movements. Thus, it is possible for a power producer to take advantage of dense PV plants networks by considering spatially distributed units as remote sensors. For low-density network, satellite-derived information observed in the vicinity of the power unit location offers an interesting alternative. To reduce the computational burden induced by this data source, feature-selection approaches are implemented. Usually, a correlation score is used to measure the dependence between lagged satellite-based time-series with the target feature (i.e. power production observations). However, this approach tends to provide redundant information (i.e. highly correlated pixels). To address this issue, we implement a minimal-Redundance Maximal-Relevance framework. Performance comparisons with state-of-the-art approaches are also performed.