In recent years, Italy and Europe have been implementing a decarbonization process in order to achieve the goal of carbon neutrality by 2050. This transition is driven by the need to generate an increasing share of energy from renewable sources. As a result, decarbonization requires a radical change in energy models, with the prevalence of distributed generation derived from clean sources, such as solar and wind, as opposed to the historical model of centralized generation from nonrenewable sources.

The impossibility of scheduling power generation from these sources turns out to be a challenge, in terms of resilience and security of the national electricity system, especially in terms of balancing demand and generation. It is therefore necessary to have production and load estimates at least one day in advance. In addition to the secure management of the system, such forecasts with a time horizon of 24÷72 hours ahead are also crucial for Day-Ahead Market trading.

Within this work, we focused on the operational production forecast of 8 wind farms, located between Sicily and Calabria, and on the electricity demand forecast for a medium-voltage line in Milan, both in aggregate and at some feeder. With regard to wind production, a case study of the years 2021/2022 was initially carried out, and then an operational chain running 72-hour forecasts with two daily runs was implemented.

In both cases, machine learning systems were experimented with such as: Random Forest, Gradient Boosting Machine and Support Vector Machine. With regard to the case study, it was found that predictions could be made with high reliability for each of the machine learning systems tested, provided a careful search for hyper-parameters was carried out. However, for the operational management of the predictions, there is a significant performance difference due to the inability to have validated data in the month prior to the day of the run and the inability to update the hyperparameters frequently. Therefore, a reasoned dataset was constructed from the Weibull distributions of the observed wind to reduce its dimensionality and computation time.

From this dataset, a set of elementary forecasts were extracted, varying machine learning model and forecasting weather models as input, for the realization of a forecasting optimizer. For load demand, on the other hand, a forecast of up to three days ahead was made, relying on machine learning models such as Random Forest and Gradient Boosting Machine, along with statistical methods such as functional Principal Component Analysis (fPCA). It was found that The fPCA method achieves, in most cases, higher correlation indices and lower error indices than the remaining methods tested.