LSTM-Based Forecasting of Danish Electricity Imbalance Price: A Comparative Study of Classical and Machine Learning Models
Translated title
LSTM-Based Forecasting of Danish Electricity Imbalance Price
Term
4. term
Education
Publication year
2025
Submitted on
2025-05-27
Pages
42
Abstract
Forecasting imbalance prices in electricity markets has become increasingly relevant due to volatility in part from renewable integration and regulatory reforms. This thesis explores the feasibility of modelling the Danish DK1 balancing market using LSTM, with the goal of enabling flexible assets to act after the clo- sure of the ID market. A stateful LSTM model was trained under cuDNN constraints utilized GPU acceleration, mixed precision training, and memory optimization to support a hyper- parameter tuning process that would other- wise have been computationally prohibitive. The tuning itself was performed using a combi- nation of Hyperband and Bayesian Optimiza- tion. The LSTM model was benchmarked against ARIMA, naive baselines, and methods such as Random Forest and XGBoost. Results showed consistent improvements in predictive accuracy after transforming the target variable through co-integration (assuming unit root) with the day ahead spot price, capturing the structural anchoring relationship and improv- ing the signal to noise ratio. Despite the com- plexity of LSTMs, simple AR models performed surprisingly well after transformation, under- lining the short term dynamics that dominate the imbalance market. Ultimately, the opti- mized LSTM model demonstrated strong per- formance in capturing extreme price spikes and directional shifts, opening for post ID op- erations. The thesis highlight the importance of aligning model design with market struc- ture and how simple models, under the right conditions can serve as an alternative.
Keywords
LSTM ; ML ; Cointegration ; Energi market ; Spot price ; Imbalance price ; mFRR ; aFRR ; Random Forrest ; XGboost ; Hyperband ; Bayesian Optimization
Documents