*3.3. Forecasting Performance*

The performance of the five day-ahead wind power forecasting models described in Section 2.1.1 is studied in this section. The neural networks models were implemented in Keras [45], using the original training procedure presented in the previous section, and the tree-based models (RF and GBDT) were implemented in Python using the Scikitlearn library [50]. The output of the ENSEMBLE model was simply coded in Python by computing the average of the outputs of the four other models. The same dataset and input features than the previous section were employed, and cross-validation was also performed for the training-evaluation procedure. Figure 5 depicts the wind power forecast obtained with the ENSEMBLE model (in red) as a function of time, as well as the actual wind power generation (in black), for a random day of the test set. One can observe that, even if the forecast error remained clearly visible, the model was most of the time able to correctly capture the time of day when the peak of wind power generation occurred, which is fundamental information for the community members for scheduling their consumption for the upcoming day.

Table 3 shows the RMSE of the five developed forecast models. We observed that the RF model was the individual model which provided the smallest forecast error in this particular application, and that the ENSEMBLE model (which was simply built by taking the average of the four other models), was still able to slightly improve the accuracy. The ENSEMBLE model was therefore selected for the operational deployment described in Section 3.4.

**Table 3.** Forecasting performance in terms of Root Mean Square Error (RMSE), for the five local day-ahead wind power forecast models (Random Forest—RF, Gradient Boosting Decision Tree—GBDT, MultiLayer Percpetron—MLP, Bi-LSTM—BLSTM, and ENSEMBLE.


**Figure 5.** Time series of wind power forecast and actuals for a day of the test set.

#### *3.4. Impact on the Consumption Behaviour of the Renewable Energy Community Members*

The two data analytics modules developed in this work were deployed operationally in the E-Cloud pilot using Mindsphere [51], a cloud-based open IoT operating system developed by Siemens. Each member of the pilot community received a personal access that he used to freely connect to a dedicated web platform using his personal computer, on his own initiative. Each member was in that way able to consult general data such as his own monthly self-consumption or self-sufficiency (see definitions below), as well as the same quantities for the whole community. It is important to mention that personal data from the other community members was hidden, for the sake of privacy. As explained in the previous sections, a common (i.e., the same for each member) day-ahead renewable generation forecast under the form of a quarter hourly time series, which was refreshed every day at 12 pm, was also made available to each member, using the methodology of Section 2.1. Each member was also able to consult a typical consumption profile for the upcoming day, representative of his past consumption at the considered time of year, according to the procedure exposed in Section 2.2. Given the preferential tariff that was in application in the community for the purchase of energy which was locally produced, we expected that the members would take advantage of the information provided by the data modules, on their own initiative, in order to adapt their consumption profiles to local generation, thereby decreasing their energy bill.

The modules became effectively operational from April 2020 to June 2020. A downscaled version of the regional solar forecast made available publicly by the TSO Elia [43] was employed for the 70 kW of solar generation installed in the community, since the absence of metered solar data in the pilot prevented the training of a local solar forecast model. The amount of installed wind generation—1.8MW—was however 20 times higher than installed PV power, which mitigated the necessity to have a very accurate solar forecasting module in this particular case.

We show the impact of the data analytics modules on the behaviour of the community members by computing the monthly self-consumption of members, i.e., the ratio between the member self-consumed energy (i.e., the member electrical energy consumption covered by the local energy which was put at his disposal) and the local energy which was put at his disposal (i.e., the portion of local generation that was allocated to him, according to predefined distribution keys mentioned in Section 3.1), during one month. This first index quantified to what extent the community generation tended to be consumed locally, where it has been produced: a self-consumption of 100% for a member meant for instance that he had consumed all the local renewable generation that was allocated to him

for the considered month. The monthly self-consumption SC*i*,*<sup>m</sup>* of member *i* during month *m* was in that way expressed as follows:

$$\text{SC}\_{i,m} = \frac{E\_{\text{self},i,m}}{\eta\_i E\_{\text{gen},m}}, \forall i \in \mathcal{Z}, m \in \mathcal{M}, \tag{4}$$

with *<sup>E</sup>*gen,*m* the total energy generated locally in the community during month *m*, *ηi* the fraction of that energy that is allocated to member *i* (constant during the whole pilot), *<sup>E</sup>*self,*i*,*<sup>m</sup>* the energy consumed by member *i* during month *m* that was covered by *ηiE*gen,*<sup>m</sup>*, and I (M) the set of community members *i* (respectively considered months *m*). It should be noted that if renewable energy was allocated to a member who did not consume it entirely, the corresponding excess of energy was not counted in the numerator of (4).

Similarly, we computed the monthly self-sufficiency of members, i.e., the ratio between the member electrical energy consumption covered by the local generation that was allocated to him and the total energy consumed by the member, again during 1 month. This index shows what part of his electricity consumption the member consumed from local resources, and by extension what part he had to purchase on the traditional retail market: a self-sufficiency of 100% means that the member was able to cover all its consumption with the local generation that was allocated to him during the considered month. Self-sufficiency of member *i* during month *m* is in that way computed as follows:

$$\text{SS}\_{i,m} = \frac{E\_{\text{self},i,m}}{E\_{\text{cons},i,m}} \; \forall i \in \mathcal{Z}, m \in \mathcal{M}\_{\prime} \tag{5}$$

with *<sup>E</sup>*cons,*i*,*<sup>m</sup>* the total energy consumed by member *i* during month *m*.

The left part of Figure 6 shows the monthly self-consumption of the 18 community members during the pilot duration, i.e., from July 2019 to June 2020. It is first very important to note that the data analytics module were effectively deployed on-site in April 2020, one month after the generalized lockdown that occurred in Belgium as a consequence of the COVID-19 crisis. The economic activity of the 18 companies involved in the community suffered in that context from a drastic reduction, which has been materialized by a significant drop of their electricity consumption, while the generation remained unchanged compared to the pre-COVID situation. This explains in the authors opinion why the self-consumption of almost all members significantly decreased starting from March 2020 to May 2020, with a progressive increase in May and June 2020, in line with the progressive removal of lockdown measures that occured in mid-May 2020 in Belgium. This effect masked unfortunately the possible positive impact of the data analytics modules on the behaviour of the community members.

The right part of Figure 6 depicts the monthly self-sufficiency of each member during the pilot duration, which should be a priori less impacted by the COVID-19 crisis since it is a ratio between two consumptions, namely the member consumption covered by local resources and the member total consumption, which are both expected to decrease in the COVID-19 situation. We observed a global increasing trend in the self-sufficiency of the community members from March to June 2020. It was however not possible to entirely attribute this positive effect to the operational availability of the developed data modules: the effect of the COVID-19 crisis on the trends in self-sufficiency could not be completely discarded, since changes in economic activity may have modified the shape of the daily consumption patterns (due to the temporary suspension of some industrial processes, etc.), which can impact the self-sufficiency as likely. Furthermore, the global increase in terms of self-sufficiency may also be attributed to a yearly seasonal effect, which is possible considering the values in July 2019, at the beginning of the pilot. The time span covered by the pilot, i.e., one year according to the special derogation granted by the Walloon regulator (CWAPE), is however not sufficient to discard or confirm that hypothesis.

**Figure 6.** Monthly self consumption (**left**) and self-sufficiency (**right**) of the 18 members of the E-Cloud pilot. The preferential tariff became effective in July 2019, which started the pilot officially, and ended late June 2020. The data analytics modules developed in this work were effectively deployed on site in April 2020, during approximately 3 months.

Finally, we show in Table 4 the relative change in terms of self-sufficiency for each member between July 2019 and June 2020, in percent. We expect in that way to compare consumer habits at almost one year interval, which can be a better indicator of possible changes in consumption patterns. For 11 out of the 18 members, it appeared that the self-sufficiency decreased, whereas it increased for six members. Again, no significant impact of the data modules on the consumption behaviour was observed (to be more conclusive, July 2020 should have been compared with July 2019, but the pilot was scheduled to end in June 2020 as explained above). The effect of the COVID-19 crisis in June 2020 could not be completely discarded as well, since the economic activity in Belgium had not recovered its pre-COVID intensity in July 2020 yet, at the time of writing.

**Table 4.** Change *SS* in self-sufficiency between July 2019 and June 2020 for each community member, in percent.


#### **4. Conclusions and Perspectives**

This work proposed energy analytics tools to inform the members of Renewable Energy Communities (RECs) of the day-ahead prospects in terms of local renewable energy generation, as well as in terms of electricity consumption profiles which are representative of the members behaviour at the considered time of the year. By doing so, the members were expected to adapt their own consumption patterns to local generation, in order to benefit of advantageous energy pricing mechanisms which prevail in a community.

A localized day-ahead wind power forecasting tool, based on state-of-the-art Machine Learning algorithms, has been developed in that way. The ENSEMBLE model, whose output is computed as the average of the outputs of four other Machine Learning models (Random Forests, Gradient Boosting Decision Trees, a MultiLayer Perceptron and a Bi-directional LSTM) has shown the best forecasting performance on the E-Cloud pilot project data. Forecasting accuracy has been further improved by automatically detecting wind power abnormal data samples and by adapting the training procedure accordingly. A procedure for generating representative electricity consumption profiles of the community members, relying on Dynamic Time Warping (a state-of-the-art Machine Learning distance employed when comparing time series), has further been implemented.

The data modules have been deployed on-site in the framework of the E-Cloud pilot project, on a REC connected to the existing Medium Voltage distribution grid in an industrial area in Belgium, and composed by 18 members (mainly Small and Medium Enterprises or SMEs) and by local generation (mainly wind power). The information provided by the data modules was freely available to the community members by connecting to a dedicated web platform on their own initiative. Global quantities, such as the monthly self-sufficiency and self-consumption of the community members, have been computed to quantify the impact of the data modules on the consumption behaviour of the community members. We were not able however to highlight significant changes in the consumer habits. It is worth mentioning though that the general lockdown that occurred in Belgium in March 2020 due to the COVID-19 crisis significantly affected the results, especially knowing that the data modules became operational for the first time in April 2020, during lockdown. Yearly seasonal effects were furthermore observed in the self-sufficiency patterns, which further masked the potential benefits of the deployed data modules. We recommend therefore to extend in accordance with the local regulator the duration of similar REC pilots to more than one year, in order to better understand these yearly seasonal effects, and to better quantify the impact on the members consumption behaviour. Furthermore, we strongly encourage researchers and industrials that will implement similar pilot RECs in the future to establish a system for monitoring the usage of the displayed information by the community members (for instance by recording the number of connections to the dedicated web platform), in order to quantify and possibly stimulate their interest in the provided tools.

As a first perspective, we intend to deploy a pilot REC for a longer time span, with an active monitoring of the members interest in the available tools, in order to confirm/infirm the hypotheses raised in this work. This is however a slow process, since temporary derogations by the local regulator are mandatory currently in Belgium for applying a community-based pricing scheme. We further aim at building another pilot REC in a residential area, in order to analyze the behaviour of domestic consumers. We finally intend to improve the accuracy of the wind power forecasting module by using turbine-level data in the model definition, and by adapting the learning procedure of tree-based algorithms to the presence of wind power abnormal data. We also intend to focus our research effort on the correct prediction of peaks of generation, since the community benefits are optimized when members shift their consumption to generation peak times, and on the recalibration of the wind power forecasting models in the flavour of [52].

**Author Contributions:** Conceptualization, Z.D.G. and D.V.; Data curation, J.B. and A.A.; Methodology, Z.D.G., J.B., A.W., P.-D.D. and J.-F.T.; Project administration, Z.D.G. and D.V.; Resources, D.V. and F.V.; Supervision, F.V.; Validation, D.V. and A.A.; Writing—original draft, Z.D.G.; Writing—review & editing, J.B., D.V., A.A., J.-F.T. and F.V.. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research has been funded by ORES. J.-F. Toubeau is supported by F.R.S./FNRS (Belgian National Fund of Scientific Research).

**Acknowledgments:** The authors would like to thank the participants of the E-Cloud project for their valuable inputs to the present study: ORES, Luminus, IDETA, Siemens, N-Side, DAPESCO.

**Conflicts of Interest:** The authors declare no conflict of interest.
