A fundamental requirement of the power grid is maintaining a balance between consumption and generation. When the balance is not fulfilled, the actors responsible for the imbalance are charged fees based on an imbalance price determined after the delivery period. Consequently, imbalance prices play a crucial role in the electricity market, and there is a need for good-quality imbalance price forecasts. 

This thesis project first focuses on probabilistic forecasting of imbalance prices using four different models: Quantile Regression, Quantile Regression Forest, a Probabilistic Neural Network, and an ARMA model.  These models were evaluated on Finnish imbalance price data of 2024, using metrics such as Pinball Loss, Continuous Ranked Probability Score, and Probability Integral Transform Histograms. A novel metric, the Relative State Performance Score, was introduced in this thesis to evaluate the prediction of imbalance direction. The evaluation demonstrated that Quantile Regression and Quantile Regression Forest had the best error metrics and the best shape of the forecast. In contrast, the ARMA model performed poorly overall but proved promising in predicting the direction of imbalance.

The second focus of the project is on a regulatory change in Finnish imbalance pricing introduced on June 12th, 2024, which caused a change in price behavior. These cases of change pose challenges for handling such a price shift from a forecasting perspective. Three different methods to handle a shift were examined through different traits using artificially generated price shifts: 1) Static Training Window, 2) Dynamic Training Window, and 3) Binary Signaling. In terms of adaption, the Dynamic Window method yielded the best results, especially for more severe shifts. By examining the models' reaction to the shift, e.g., by overconfident predictions, the results showed different properties for the underlying models, but overall, the Binary Signaling method avoided overconfident predictions.