In this work, demand and renewable energy production forecasting problems are treated. The main goal of this work is the prediction’s accuracy improvement using machine learning methods and time series analysis too.
In the case of demand forecasting problem we have a five years dataset of real measurement of ambient factor obtained in Chieti town in Italy, energy price and demand. We say ambient factor to indicate external temperature, dry bulb temperature, humidity ecc. The sampling period is 30 minutes.
The goal of this problem is having good accuracy in Short-Term-Load-Forecasting (STLF), i.e. 30 minutes ahead forecasting. STLF problem are very useful in real cases. Just for an example, in Italy TERNA group, which is the energy dispatcher, should have short term good prediction’s energy demand to better balancing electric network.
Many strategies based on multiplicative ARIMA model or Support-Vector-Machine for regression problems are led to resolve the problem. All methods are then compared in terms of prediction results on a test set.
For the second problem (renewable energy production forecasting) we have five datasets corresponding to five wind farms proprieties of an important italian enterprise. These datasets contain metering data corresponding to a grid points built around each wind park. These metering data are obtained every day at midnight by a software which generetes data for the same and next day. The enterprise uses next day metering data to estimate wind park power production and it makes sell trusts with this estimation. The sampling period is one hour.
The main goal of this problem is having good accuracy in energy production forecasting so that the enterprise won’t pay higher penalty. We used models based on Support-Vector-Machine for regression problem and an important previous step of feature selection. An overview of some important feature selection techniques such as RRelieff algorithm or LASSO technique is discussed in this work. The goal of our feature selection is finding the best points of the grid built around each park.
We say “best” point to indicate grid point which has good power prediction. We build a pipeline model for feature selection which contains more block based on feature-ranking and wrapper methods.
Our pipeline is then compared to enterprises’s feature selection technique in term of prediction results on a test set.
Finally we considered unbalancement problem. Sometimes higher or lower penalties are given to enterprise when it underestimates or overestimate energy production. These penalties also depend on zonal unbalancement. We built a simple Generalized Stochastic Petri Net (GSPN) to model unbalancement problem so that the enterprise can decide when under/overestimating is convenient in term of penalty.
Advisor: Marco Sciandrone and Fabio Schoen
Candidate: Guido Cocchi
Laurea: Magistrale in Ingegneria Informatica