Comparative Evaluation of Statistical, Machine Learning and Deep Learning Models for DayAhead Electricity Price Forecasting: A Case Study of the Polish Market

Electricity Price Forecasting (EPF) for the Polish Day-Ahead market (TGE/POLPX). In the context of rising renewables, expanding storage and increasingly dynamic tariffs, reliable day-ahead electricity price forecasting has become essential for trading, system balancing and investment planning. Forecast accuracy directly affects balancing costs, forward-contract profitability and storage returns, and the effectiveness of predictive models depends on market characteristics. Despite extensive research for major European markets, a comprehensive comparative study for Poland (TGE/POLPX) has been lacking. This thesis addresses that gap by evaluating representative approaches across statistical, machine-learning and deep-learning families to determine which class is most suitable under Polish market conditions. Data scope: Eight years of hourly prices (2016–2023) for the Polish day-ahead market, enriched with a uniform set of calendar and fundamental features; all models receive identical inputs. Model families: Ten representative methods: statistical (ETS, SARIMA, SARIMAX, TBATS, Prophet), machine learning (SVR, Random Forest, XGBoost) and deep learning (LSTM, TCN). Validation protocol: A daily-shifted rolling-origin design with an ≈30-day training window and 24-hour forecast horizon; consistent training/tuning and common evaluation metrics ensure a like-for-like comparison. Materials: The thesis documents the full pipeline and design choices. An open replication package (code and data pointers) will be released via the associated GitHub repository.