Improved Cleaning Event Detection Methodology Including Partial Cleaning by Wind Applied to Different PV-SCADA Datasets for Soiling Loss Estimation

Soiling on photovoltaic (PV) systems reduces operating efficiency, requiring frequent cleaning and maintenance. To quantify soiling loss directly from the PV performance time-series data, an essential step is to detect cleaning events accurately. However, analyzing the time-series data of utility-scale PV plants comes with its own set of challenges. Sensors not calibrated or maintained correctly, missing or inconsistent time stamps, defective components (PV modules or inverters), and incomplete maintenance logs can affect the process. Due to these reasons, daily variations in the PV performance ratio time-series data tend to be high, making it noisy. To overcome the above issue, a modified 1-sigma Hampel filter is used in this work, which aims to remove noise from the PV performance time-series data. The Hampel filter uses a rolling window and median absolute deviation to detect outliers in time-series data. Compared with the recent state-of-the-art filters used in time-series analysis, the proposed filter can effectively deal with noise with a Gaussian or non-Gaussian distribution, is easy to implement, and is resilient to local outliers in the data. The filter is applied to three PV datasets of different capacities to detect cleaning events accurately. In addition, using SCADA data analysis, we also discuss the role of wind speed and relative humidity in PV modules' cleaning, which was reported to result in panel cleaning under specific controlled experiments. Through the comprehensive analysis of SCADA data and accompanying metainformation, we show that our algorithm can detect cleaning under high wind speed and low relative humidity.