SPLTRAK Abstract Submission
Real-time Prediction Algorithms to Detect Clouds and Forecast Photovoltaic System Performance
Mughal Maqsood, Muhammad Uddin, Habeebullah Adua, Evan Sauter, Stephen Natale, Timothy Lewis, Jonathan G. Ferreira
, , ,

The utility industry needs an immediate solution to the sharp short-term changes in PV energy and voltage flicker. When a cloud shadow covers a PV system, PV losses can be up to 25%, and as a result, the energy supply can fluctuate below the actual demand. Using line voltage regulators, switched capacitor banks, and backup generators to alleviate the energy fluctuations is inefficient because these devices need to operate more frequently than usual. Switching from solar to backup power is not instantaneous, leading to flickering moments and equipment malfunction. Cloud studies using satellite data are only suitable for observing wide-area clouds while predicting local and real-short-time information about clouds is challenging. Machine learning techniques are inconsistent because model training depends on the existing data, which varies for different PV sites.   This work introduces a frugal innovation of a CMVS model that can compute cloud motion parameters (speed, direction, size) due to changes in irradiance from the cloud cover in the real-time frequency of 10 kHz. The model uses nine-light sensors groups to form a cluster and will be installed close to the PV site to monitor PV system energy and forecast the system's performance before the cloud covers reach the PV site. The model will trigger the backup power supply until it restores the optimal operation when the forecast predicts a fall in PV outputs below a certain level. The system also features a smart design for IoT applications with scalable deployment. The model has the potential to help grid operators mitigate the effects of PV power variability on grid planning and operations.