Recent advances in deep learning have significantly improved the accuracy of timeseries forecasting, the task of predicting how values change over time based on historical data. Time-series forecasting models are widely used for decision-making in domains such as building control, heating systems, and financial trading. However, the vast majority of state-of-the-art models still provide only a single forecast without indicating model confidence, so end-users may end up acting on unreliable predictions when using these models during decision making. This thesis proposes a solution to this problem by introducing selective time-series forecasting, where the timeseries forecasting model estimates its confidence and uses it to select or reject any forecast. This allows end-users to trade off how often the forecasting model outputs a forecast for an overall lower forecasting error of the selected forecasts, which is especially important when inaccurate predictions have significant penalties. The thesis also applies these ideas to heat pump flexibility, where time-series forecasting models are used to decide how heat pumps can adjust power consumption while keeping indoor temperature within user-defined comfort bounds. By rejecting uncertain forecasts, the methods help reduce overestimation of flexibility and lower the risk of financially expensive errors. Finally, the thesis extends selective forecasting to time-series foundation models through fine-tuning, improving forecasting performance on data unseen during training with only a small increase in additional parameters and computational over-head.

After the defense there will be a small reception in cluster 3. All interested parties are welcome.