LSTM-Based Forecasting of Danish Electricity Imbalance Price: A Comparative Study of Classical and Machine Learning Models

LSTM-Based Forecasting of Danish Electricity Imbalance Price

Christiansen, Christian Taulbjerg Gørup

4. term

Mathematics-Economics, Master

2025

2025-05-28

42

Ubalancepriser i elmarkedet er blevet mere volatile, bl.a. på grund af mere vedvarende energi og regulatoriske ændringer. Denne afhandling undersøger, om ubalancepriserne i det danske DK1-balancemarked kan modelleres med en LSTM (en type neuralt netværk, der lærer tidslige mønstre), så fleksible aktiver kan agere efter ID-markedet (intradagmarkedet) er lukket. En stateful LSTM blev trænet med GPU-acceleration og cuDNN-kompatible indstillinger samt mixed precision og hukommelsesoptimeringer for at gøre omfattende hyperparametertuning mulig. Tuningprocessen kombinerede Hyperband og bayesiansk optimering. LSTM’en blev sammenlignet med ARIMA, simple naive baselines og træbaserede metoder som Random Forest og XGBoost. En central observation er, at prognosen bliver mere præcis, når målet transformeres via kointegration med day-ahead-spotprisen (under antagelse af en enhedsrod). Denne transformation indfanger den strukturelle forankring mellem ubalance- og day-ahead-priser og forbedrer signal-støj-forholdet. Efter dette trin klarede selv simple autoregressive (AR) modeller sig overraskende godt, hvilket peger på, at kortsigtede dynamikker dominerer ubalancemarkedet. Den optimerede LSTM var dog særlig stærk til at fange ekstreme prisspidser og retningsskift, hvilket er vigtigt for drift efter ID-markedets lukning. Samlet viser arbejdet betydningen af at tilpasse modeldesign til markedsstrukturen og, at simple modeller under de rette betingelser kan være et praktisk alternativ.

Imbalance prices in electricity markets have become more volatile, partly due to renewable integration and regulatory changes. This thesis examines whether an LSTM (a neural network that learns patterns over time) can model imbalance prices in the Danish DK1 balancing market, so that flexible assets can act after the intraday (ID) market closes. A stateful LSTM was trained with GPU acceleration and cuDNN-compatible settings, using mixed precision and memory optimizations to make extensive hyperparameter tuning feasible. The tuning combined Hyperband and Bayesian optimization. The LSTM was benchmarked against ARIMA, simple naive baselines, and tree-based methods such as Random Forest and XGBoost. A key finding is that forecasting accuracy improves when the target is transformed via cointegration with the day-ahead spot price (assuming a unit root). This transformation captures the structural anchoring between imbalance and day-ahead prices and improves the signal-to-noise ratio. After this step, even simple autoregressive (AR) models performed surprisingly well, indicating that short-term dynamics dominate the imbalance market. The optimized LSTM nevertheless excelled at capturing extreme price spikes and directional shifts, which is critical for operating after the ID market closes. Overall, the work highlights the importance of aligning model design with market structure and shows that, under the right conditions, simple models can be a practical alternative.

[This summary has been rewritten with the help of AI based on the project's original abstract]

LSTM ; ML ; Cointegration ; Energi market ; Spot price ; Imbalance price ; mFRR ; aFRR ; Random Forrest ; XGboost ; Hyperband ; Bayesian Optimization

Download
View record in AAU Student Projects