Evaluation of Various Deep Learning Models for Short-Term Solar Forecasting in the Arctic using a Distributed Sensor Network

Abstract

The solar photovoltaic (PV) power generation industry has experienced substantial, ongoing growth over the past decades as a clean, cost-effective energy source. As electric grids use ever-larger proportions of solar PV, the technology's inherent variability-primarily due to clouds-poses a challenge to maintaining grid stability. This is especially true for geographically dense, electrically isolated grids common in rural locations which must maintain substantial reserve generation capacity to account for sudden swings in PV power production. Short-term solar PV forecasting emerges as a solution, allowing excess generation to be kept offline until needed, reducing fuel costs and emissions. Recent studies have utilized networks of light sensors deployed around PV arrays which can preemptively detect incoming fluctuations in light levels caused by clouds. This research examines the potential of such a sensor network in providing short-term forecasting for a 575-kW solar PV array in the arctic community of Kotzebue, Alaska. Data from sensors deployed around the array were transformed into a forecast at a 2-minute time horizon using either long short-term memory (LSTM) or gated recurrent unit (GRU) as base models augmented with various combinations of 1-dimensional convolutional (Conv1D) and fully connected (Dense) model layers. These models were evaluated using a novel combination of statistical and event-based error metrics, including Precision, Recall, and F β . It was found that GRU-based models generally outperformed their LSTM-based counterparts along statistical error metrics while showing lower relative event-based forecasting ability. This research demonstrates the potential efficacy of a novel combination of LSTM/GRU-based deep learning models and a distributed sensor network when forecasting the power generation of an actual solar PV array. Performance across the eight evaluated model combinations was mostly comparable to similar methods in the literature and is expected to improve with additional training data.