Renewable Energy Communities Digital Twin (RECDT)

This project was conceived after the project “Planning Renewable Energy Communities 
(REC): the perspective of Municipalities and Small and Medium Enterprises 
(SME)” (PLANREC), that was as well part of the ERIGrid 2.0 Lab Access programme. In this 
previous project we developed a tool for the forecast of energy performance and metrics of 
Renewable Energy Communities.  
The goal of this new project is to investigate whether it is possible to obtain similar results by 
relying exclusively on socio-economic data of the REC, excluding historical consumption data 
of individual households from the necessary prediction data. In most proposed approaches, the 
short or long-term history of energy consumption is used to predict future energy load de
mands, but the historical load data needed for this kind of model is not always available. With 
our work we try to overcome the possible lack of this data and enable accurate REC energy 
forecasts even when the historical data are not available. 
Even though the proposed model does not require historical load data to make its predictions, 
it still needs historical load data for the training phase.  We searched for public sources that 
included both electricity load and socioeconomic data for the training and testing of the model. 
We found many datasets, but the most complete for our purposes was the IDEAL dataset 
(IDEAL). We created a database in pgAdmin by reworking the IDEAL dataset for our purposes, 
in which we included the weather data and the location of the households. 
It is important to notice that, since we do not have any historical data, our task is a regression 
problem, not a time-series forecasting problem. We tested some regression models: - 
RandomForestRegressor; - - 
ExtraTreesRegressor; 
XGBoost Regressor. 
From the grid searches that we conducted on these models, we observed that the best per
forming model was the ExtraTreesRegressor. An important thing to remember is that even 
though in the best case the ExtraTrees score was slightly better than the RandomForest and 
XGBoost best score, the results of the three models were very similar.  
What really made the difference in the model performance was the introduction of synthetic 
feature related to the load of the households: - 
Mean_daily_min_load: average of the daily minimum consumption for each household; - - 
Change_point_down: indicates whether the household is in stand-by (only basic con
sumption) or if there are other consumption that can be attributed to the presence of 
people; 
Change_point_up: indicates the presence of consumption peak. 
We validated the performance of the model with R-squared score and the final result was a 
score of 0.82. We also tested the model with 24-hour forecasts for some specific households, 
achieving a prediction of the consumption patterns with an acceptable degree of accuracy.