In countries such as India with low grid prices, energy firms are offering competitive PPA tariffs for solar farms. Given lower mar- gins in operating these farms, there is great sensitivity to panels under-performing. To detect under-performance, condition- monitoring methods compare generated power with an ideal yield calculated for localised weather. Applying such methods to a 1.2MW farm with 6 different PV technologies over 3 years, we ob- served prediction errors large enough to mask under-performance. To reduce this error, we explicitly modelled the Maximum Power Point Tracker (MPPT) in a two-step prediction method. In doing so, we combine a regression method on weather data with physical modelling of IV-characteristics of panels, resulting in an average reduction in error by 16%.