DGImNet: A deep learning model for photovoltaic soiling loss estimation

Deep learning models have recently been introduced to photovoltaic (PV) soiling loss estimation tasks. Most PV soiling loss (PVSL) estimation models are based on a single image and the environmental factors at a specific time point, while the temporal characteristic of environmental factors is less utilized. DGImNet, a PVSL estimation model utilizing both PV panel images and time series environmental factors (TSEFs), is proposed. DGImNet takes a PV panel image and TSEFs of 50 continuous time points to estimate the PVSL. The TSEFs are processed by gate recurrent units to produce a 96D feature, while the image is extracted to generate another 96D feature by a set of analysis units. The multi-modal features are fused to yield estimation results. It is proved that the exploitation of TSEFs is beneficial to improve PVSL prediction performance, and the engagement of compact bilinear pooling is useful for better fusion of image and TSEF features. Cooperating with a real-time data collection system, the proposed model is able to run on edge computing devices and be employed for real-time PVSL estimation tasks in actual PV power stations.