Uncertainty Quantification in Fuel Planning Based on Historical Forecast Deviations : Modeling the gap between planned and actual fuel consumption for improved decision support in fuel procurement

Sammanfattning

EFO AB is responsible for sourcing and coordinating fuel deliveries toseveral Swedish energy producers, including Mälarenergi. At Mälarenergi,the Energy Optima 3 (EO3) optimization model is used to generate productionplans for individual boilers. These plans form part of the basis for fuel-ordering decisions. However, the realized boiler operation may deviate fromthe optimized plan, creating uncertainty for EFO’s sourcing team. Suchdeviations can lead to excess inventory, storage pressure, fuel degradation,or short-notice deliveries.This thesis investigates whether historical deviations between EO3-basedoptimization outputs and realized boiler operation can be used to quantifyuncertainty in fuel planning. The analysis is based on historical planning andoperational data for three boilers, P5, P6 and P7, at the Mälarenergi plant inVästerås. A two-part modeling structure is used: First, logistic regression isapplied to estimate the probability that a planned boiler run does not occur.Second, multiple linear regression is used to model the signed deviationbetween planned and realized production, conditional on the boiler actuallyoperating.The logistic regression model shows strong classification performance,with a ROC-AUC of 0.959 and a recall of 0.968. The results indicate thatboiler P7 has a substantially higher probability of missed operation than theother boilers, and that higher temperatures are associated with a higher missed-run probability. The multiple linear regression model provides additionalinsight into the size and direction of deviations when operation occurs. Thebest subset model includes planned MWh, spot price and boiler identity, butits predictive performance is more limited, with a mean absolute error ofapproximately 501 MWh.The results suggest that EO3 outputs should not be interpreted asdeterministic production plans, but as uncertain planning signals whosereliability varies across boilers and operating conditions. The logisticregression model is the most immediately useful component, as it can serveas a warning layer for planned runs with elevated uncertainty. However, thecurrent models rely partly on realized temperature and spot price, which arenot known at the time of sourcing decisions. Future work should thereforeincorporate decision-time variables, such as weather forecasts, spot priceforecasts and EO3 input assumptions, before the approach can be usedoperationally.