This blog entry is available on Medium
Herein are two examples about how a small gap of understanding the Solar PV Systems Modeling & Analytics could lead to misleading conclusions:
1) There are several studies on solar forecast where the forecast models are trained with measured data of weather conditions collected from ground weather stations, and these models are tested against the unseen part of measured data. So, in the real-world applications, those models are failing to obtain as good performance as they did in their test stage! ..Why? Because, in case of real-world applications those models now use data of similar quantities as they trained with but the data are now future forecasts rather than past measurements. In other words, instead of using measured data from ground stations, now those models must use future weather forecast from NWP or image projections from sky imagery devices and satellites. Therefore. from a practical point of view, the forecast models should also be trained with weather forecasts that recorded in past rather than ground measurements, and hence they can be implemented for solar forecasts with reasonable accuracy as well as properly quantifying the deterministic and probabilistic (uncertainty) errors associated with their forecasts.
2) Some folks (with pure AI, ML backgrounds) are trying to test the forecasts with economic dispatch (caring only about the cost) to convince the electric operators about the value of their forecast models. Meanwhile, for electric power systems, the forecasts should be tested against the security constrained economic dispatch or the optimal power flow should be run with the forecasts to check out the cost as well as the security/reliability constraints are not violated (constraints such as, generators, transmission lines, and node voltages rating and limits).