Prognostisering av aggregerad elanvändning och produktion för portföljer med balansansvar : En jämförande studie med Transformer-, XGBoost- och LSTM-modeller

This thesis addresses the challenge of accurately forecasting day-ahead electricity consumption and production for bidding purposes in Nord Pool’s Day-Ahead Market. In this highly interconnected environment, even minor forecast errors can result in suboptimal bidding strategies and elevated imbalance costs for Balance Responsible Parties in the electricity system, highlighting the importance of robust and reliable prediction methods. In this study, the performance of Autoregressive Integrated Moving Average (ARIMA) as a baseline model is compared against more advanced approaches: Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM) networks, and the Temporal Fusion Transformer (TFT). These models are evaluated based on their accuracy—measured through MAE, nMAE, MAPE, and RMSE—as well as interpretability, with the goal of identifying the most effective approach for day-ahead consumption and production forecasting. The results indicate that the Temporal Fusion Transformer demonstrates strong potential, achieving the lowest error metrics across the board among the models evaluated. Moreover, by leveraging Quantile Regression, TFT natively produces probabilistic prediction intervals, providing valuable uncertainty estimates without the need for external post-processing—a critical advantage for operational bidding under uncertainty. For the more challenging production case, a hybrid approach combining an XGBoost-based classifier with the TFT model further improved performance. The classifier outputs a probability of production per hour, which is then passed as a known real-valued input to the TFT model. This setup enhances the model’s ability to anticipate sparse and volatile production events, especially under high variance and zero-heavy regimes. Further, a simple theoretical framework for bidding optimization is proposed leveraging linear programming, in order to try to maximize revenue and minimize cost in the day-ahead market while staying risk-averse. By offering a clearer picture of future energy supply and demand, this forecasting framework can support more effective bidding strategies, reduce imbalance risks, and enhance overall market efficiency. This thesis makes a valuable contribution to the growing body of work in data-driven energy forecasting and lays the foundation for future research in hybrid modeling and real-time energy operations.