Techno-economic assessment of the realisable economic value and value capture gap of probabilistic PV forecasting in PV-battery microgrid dispatch

Photovoltaic (PV)-battery microgrids require short-horizon forecasts for storage dispatch, yet statistical accuracy alone cannot establish value because storage state, grid recourse, curtailment, and shortage cost shape outcomes. This study quantifies the realisable economic value and value capture gap of probabilistic PV forecasting using 16 months of data from the Desert Knowledge Precinct microgrid in Alice Springs, Australia. Four lightweight point and quantile models are evaluated in a two-stage linear dispatch framework with perfect forecast bounds, persistence baselines, fixed commitment recourse tests, and islanded stress tests. In the central grid-connected case, perfect forecasting reduces operating cost by Australian dollars (AUD) 5223, or 9.8%, relative to persistence. k-nearest neighbours (kNN) Q 0.7 provides the highest tested central value, saving AUD 1,124 and capturing 21.5% of the perfect forecast value. Under a 150 kW import limit and 20 AUD/kWh value of lost load, it reduces persistence cost by 35.8%, close to the 40.3% perfect forecast bound. Under islanding, Light Gradient Boosting Machine Q 0.1 is the most robust low quantile and slightly outperforms persistence. Thus, translating forecast skill into economic value is conditional on model family, quantile choice, storage state, recourse capacity, and reliability cost, and should be assessed through dispatch outcomes.