*1.2. Related Work*

Many challenges related to the modeling of RECs are still investigated in the literature, which mainly deal with the optimal operation (e.g., how should we allocate in day-ahead energy resources among members in a community to fulfill a given objective?) and sizing (how should we dimension renewable generation, storage, etc. in a community?) of the communities (see references [6–11]

exposed above). More particularly, the multi-agent nature of the underlying optimization problems has driven an increasing attention of the researchers towards game theoretical models for studying the economic equilibria that can appear inside the communities [11,19,20]. Some authors are focusing on the other hand on regulatory aspects related to RECs: these communities consist indeed in a new market design which can play a role at the macroeconomic scale in case of a general adoption. Authors in [21] use cooperative game theory to show for instance that inadequate grid tariffs may lead to an excess adoption of the model, with a potential snowball effect.

Optimally allocating resources in day-ahead in a community requires however to be informed with accurate prospects in terms of local injections (renewable energy production) occurring in the community, with a small time granularity (e.g., quarter hourly). The role of demand (i.e., electricity consumption) response for better matching the generation, in the context of RECs and more generally in electrical power systems, is furthermore well-kown and heavily investigated in the literature, through e.g., the direct control of appliances (see e.g., [22]), or appropriate ex ante recommendations on the consumption behavior of end-users, provided possibly by optimization routines driven by economic signals (see e.g., [23]).

More particularly, data analytics techniques, and especially Machine Learning, can play an important role in better anticipating the generation and demand primitives in communities. The 1 h ahead forecasts of the electricity consumption are for instance performed in [24] using neural networks, in order to support a fuzzy-logic based controller which implements the resource matching in rural communities. Authors in [25] developed a Markov Chain for forecasting a day ahead the aggregated solar generation surplus and residual load in a community comprising storage. A Long Short Term memory network is proposed in [26] to forecast in day-ahead the energy demand in a whole P2P community. Other researchers try on the other hand to avoid the complexity of forecasting models by developing online optimization methods [27,28]. Finally, some studies leverage data analytics for improving the sizing of the communities, such as [29], in which a load profile generator based on Self Organized Maps (SOMs) is proposed.

## *1.3. Objectives and Contributions*

In this paper, we focus on the day-ahead forecasting of time series of local wind power generation in a community, whereas most of the literature studies communities with solar generation only, and on the modeling of the electricity consumption of the individual community members, whereas most of references focus on the consumption quantities aggregated at the community level. We develop data analytics modules, relying on state-of-the-art Machine Learning models, which are expected to help the community members to adapt their consumption profiles to the local renewable energy generation, thereby improving the local coordination. More particularly:


18 Small and Medium Enterprises (SMEs), who had the possibility to freely access the results of the developed data modules by connecting to a dedicated web platform,

5. we quantify the impact of the modules on the operation of the REC (forecasting performance of the developed models, and behaviour of the community members via the evolution of their self-sufficiency and self-consumption during the pilot).

The paper is organized as follows. Section 2 describes the developed data analytics models, with an emphasis on local wind power forecasting in Section 2.1 and on the generation of representative electricity consumption profiles in Section 2.2. Section 3 first describes the pilot REC on which the developed modules are applied (Section 3.1), focuses then on the performance of the wind power forecasting module (Sections 3.2 and 3.3), and finally quantifies the impact of the data analytics modules on the behaviour of the community members and on the operation of the REC, by analyzing the evolution of the community self-consumption and self-sufficiency (Section 3.4).
