Using Siamese Networks to Detect Shading on the Edge of Solar Farms

Abstract

Solar power is one of the most promising sources of green power for future cities. However, real-time anomaly detection remains a challenge. Internet of Things (IoT) is an effective platform for real-time monitoring of large-scale solar farms. Using low-cost edge devices such as the Raspberry Pi (RPI), it is possible to not only read power and irradiance values from in-situ sensors, but to also apply machine learning and deep learning algorithms for real-time analysis and for detecting anomalous behaviors. This paper presents the design and implementation of an edge analytics application that uses RPI as an edge device. The Isolation Forest algorithm was first used to detect shading anomalies. A Siamese neural network was then trained to create a latent-space mapping. An anomaly detection model based on the latent space and a neural network and kNN was developed. These models could detect shading anomalies with an F1-Score of 0.94. Embedded variants of the model based on TensorFlow Lite and TensorRT were evaluated to service a large number of solar panels at 1Hz. The results are that a single RPI could do parallel anomaly detection of 512 solar panels at 1 Hz with 0% failures. The TensorRT variant consumed more resources than the TensorFlow Lite implementation, but the maximum CPU utilization remained below 75%.