Abstract

Energy time series sit at the heart of monitoring, forecasting, and control in smart buildings and power systems. In real deployments, however, the data rarely match the assumptions of clean and stationary signals. Sensor faults, missing readings, and communication interruptions introduce abnormal observations, while demand patterns drift with weather, occupancy, and equipment operation. These issues can quietly erode predictive performance and, when uncertainty is not made explicit, can also mask risk in decision-making. This thesis improves reliability through two connected contributions. First, it investigates unsupervised anomaly detection for energy consumption signals. The approach learns typical temporal behavior from clean data and flags departures without relying on labeled anomalies. Beyond detection metrics, the contribution is evaluated through practical impact on forecasting: we quantify howcorrecting anomalous segments changes downstream prediction errors and whether it stabilizes performance under data corruption. The second line of work focuses on probabilistic load forecasting. Rather than producing a single point forecast, the aim is to produce a predictive distribution that accounts for variability and uncertainty. The evaluation pairs accuracy measures with an assessment of calibration and how well uncertainty estimates hold up under changing conditions. In combination, these contributions emphasize that performance in real deployments depends on both signal integrity and uncertainty quantification. Anomaly handling reduces the impact of corrupted observations, while probabilistic forecasting better supports operation under variability and drift