• deep learning
• forecasting
• italian market
• machine learning
• pinn
• price
• transformer

In a complex and evolving energy market, forecasting electricity prices
and volumes is essential for optimizing trading strategies, minimizing
waste, and meeting rising electricity demand. This thesis focuses on
improving electricity price and volume forecasting within the Italian
market by accounting for its zonal structure, the influence of renewable
energy sources, and interactions with foreign markets.
The primary aim of this thesis is to analyze the unique characteristics
of zonal prices in the Italian electricity market and to develop machine
learning models capable of accurately predicting these values, with
a particular focus on attention-based models, such as Transformers.
The work begins with an in-depth examination of the structure and
functioning of the Italian electricity market. Chapter 1 provides a
detailed description of the key variables influencing market trends
over time. The volatile nature of energy prices, driven by different
seasonality, highlights the necessity for robust forecasting models to
enable market participants to make informed and effective decisions.
Chapters 2 and 3 delve into the state of the art of Electricity Price
Forecasting (EPF), introducing various neural network models tailored
for this task. These chapters also provide an overview of the mathemat-
ical foundations underlying these models. Specifically, attention-based
neural networks, such as Transformers, are explored in depth to con-
tribute to the ongoing debate on their effectiveness in time series
forecasting (TSF). These models are applied to the specific case of each
market zone in the Italian system.
Furthermore, an ensemble approach is proposed, integrating the
best-performing neural network models with a physics-informed neu-
ral network (PINN). This approach incorporates domain-specific con-
straints into the loss function, leveraging known physical properties
of the system to improve prediction accuracy. Notably, constraints
derived from market resolution conditions are introduced, enhancing
the realism and reliability of the predictions.
Finally, Chapter 6 presents a comparative analysis of the differ-
ent modeling approaches, highlighting their respective strengths and
weaknesses, such as interpretability, computational efficiency, and
accuracy in extreme scenarios, more commonly denoted as spikes.
Chapter 7 concludes with a discussion of potential improvements and
future research directions.
The findings underscore the importance of combining advanced
neural network architectures with domain-specific constraints to pro-
duce more accurate and reliable forecasts in the dynamic and complex
context of the electricity market.