Forecasts judged by the decisions they enable

My doctoral thesis, Advanced Photovoltaic and Load Power Forecasting Methods for Optimal Microgrid Scheduling, studies how forecasting and optimization should be designed together. Conventional practice trains forecasting models for statistical accuracy and hands their output to a scheduler. I show that this separation leaves value on the table, and develop forecasting methods that are trained on the cost of the decisions they inform.

The work runs from method to deployment: cost-aware and decision-focused forecasting, guarantees for models that keep learning while the system operates, coordination of several forecasters within one energy management system, risk-aware training for the days that matter most, and diagnostic tools that tell an operator which forecast errors are actually costly. Every method is validated on a real laboratory microgrid with photovoltaic generation, battery storage and electric vehicle charging, and within European research projects.