As the energy sector transitions toward renewable energy integration, balancing electricity production and consumption has become a critical challenge. Imbalance prices‚ reflecting the cost of rectifying grid discrepancies‚ play a central role in power market dynamics and trading strategies. Accurate forecasting of imbalance prices can help power companies limit financial risks and guide trading decisions. This thesis investigates machine learning-driven time series forecasting, employing the Extreme Gradient Boosting (XGBoost) algorithm to predict imbalance prices in Sweden’s SE2 bidding area.
The research encompasses data gathering, preprocessing, feature engineering, and evalu-ation against statistical baselines such as ARIMA and naive benchmarks. The proposed XGBoost model outperforms the naive model by over 25% and the ARIMA model by over 20% across four error metrics: MAE, MdAE, P90AE and RMSE, indicating a measurable improvement in predictive performance.