Following the decline in inverter production by inverter, a multimodal model anticipates cloud passage and advances 14 reference methods, including overcast or changing skies.
Integrating more photovoltaics into the grid requires knowing, to the nearest hour, how much the panels will produce, whereas the passage of a cloud can cause power to drop in a few minutes. The authors of IMPACT propose a physics-guided deep learning model that provides probabilistic forecasts of 10 to 110 minutes. Its originality: exploiting the plant's inverters as a network of sensors, whose successive decreases reveal the progression and heterogeneity of cloud cover.
Four combined sources of information
The model processes separately the plant's production history, the spatial dynamics of the clouds as seen by the inverters (via a graph whose links are learned), the solar geometry known in advance and the current weather state, before merging the whole thing. Rather than predicting production from scratch, it corrects a persistence prediction, and its confidence intervals are continuously recalibrated.
Compared to 14 reference methods, IMPACT obtains the best prediction skill score (0.295, i.e. nearly 0.1 more than the second) and maintains a clear advantage by overcast (0.198) or variable (0.255). Tests by removing each component in turn show that they all contribute to performance.
Read the full article on the original source
Go to the source →