Energy Price Forecasting in Italy

Summary

The aim of this thesis is to investigate the effectiveness of Machine Learning algorithms in forecasting electricity prices, with a particular emphasis on the Italian wholesale electricity market (IPEX), a key reference point in South Europe. Utilizing domain knowledge, an extensive dataset comprising 168 variables was compiled. The study encompasses the application of various conventional machine learning techniques and artificial neural networks, employing prominent Python libraries like scikit-learn and keras. Electricity markets are constantly changing, which necessitates the updating of our algorithms’ training datasets. Based on our experiments, we found that the ideal training dataset should be rolling and include only the last 15 days. Predictions should be made only for the next day and not for a longer period, as the error increases with the length of the forecast interval.

Additionally, for the validation of the results, nested cross-validation was used instead of simple cross-validation because we have time-series data and we want to avoid data leakage. The nested cross-validation procedure provides an almost unbiased estimate of the true error. As we progress from basic to more advanced methodologies, there is a clear trend of enhanced performance. We observed a reduction in the Mean Absolute Percentage Error (MAPE) from around 20% to 5%, a testament to the power of artificial neural networks in accurately modelling the relationship between input factors and the predicted prices. Additionally, a sensitivity analysis was conducted to assess the influence of specialized knowledge on the results, which underscored the vital role that each feature plays in bolstering the algorithms’ effectiveness.