Nordic electricity market price forecasting based on deep learning : integrating transformer networks and external features for enhanced accuracy

Abstract

The Nordic electricity market is highly dependent on renewable energy such as hydropower and wind power, resulting in violent price fluctuations and it is difficult to make accurate predictions. This thesis develops a robust forecasting framework based on deep learning and integrated learning (XGBoost) to predict short-term electricity prices in the Nordic market. The model integrates historical price data and time characteristics and uses a 168-hour retrace window to predict the price in the next 24 hours. We used a two-year Finnish data set (2022-2023) to evaluate the performance of the model and calculate the root mean square error (RMSE). The proposed XGBoost model achieves the RMSE of 75.53 Euros/megawatt-hour on the test set in 2023, which is better than the continuity model (105.40 Euros/megawatt-hour), the SARIMA model (92.15 Euros/megawatt-hour) and the LSTM model (85.60 Euros/megawatt-hour). The deviation correction mechanism calculates the average deviation between the forecast value and the actual value in the rolling window, so that the forecast value is consistent with the actual market peak.

The model performed particularly well in the period of market stability; for example, after the full operation of the Olkiluoto 3 nuclear reactor, the prediction error was significantly reduced compared with the previous period. However, extreme "black swan" events - such as unprecedented bidding errors and power grid anomalies in November 2023 - are still challenging because these events go beyond the historical model learned by the model. The main contribution of this work is to provide a practical and lightweight forecasting tool for energy companies, grid operators and policymakers. Future research will focus on integrating multi-source exogenous variables, including temperature, wind speed and nuclear power plant shutdown indicators and explore a hybrid deep learning architecture that combines Transformer-based feature extraction with XGBoost regression. These extensions aim to improve the accuracy of prediction, enhance the robustness of the model under extreme conditions, and extend the scope of application of the model to other renewable energy power markets outside Northern Europe.