Anomaly Detection in Renewable Energy Big Data Using Deep Learning

Introduction

Renewable Energy (RE) attracts countries worldwide for many potentials and benefits. Thus, Saudi Arabia has increased its focus on including renewable resources in the national energy mix. It is currently developing and investing in many energy systems. The primary reasons are contributing to its climate obligations and diversifying its economy from being fossil fuel-dependent (Barhoumi et al., 2020).

A significant challenge for the global energy supply is the large integration of RE sources, such as solar energy, into existing or future energy supply structures. An electrical operator must maintain a precise balance of electricity production and consumption. Additionally, the operator frequently needs help maintaining this balance with conventional and controllable energy production systems, particularly in small or non-interconnected electrical grids like islands. Moreover, the electrical system's reliability depends on its ability to maintain quality and continuity of service to customers. Then, the complexity of the system management across multiple time horizons, considering RE's erratic behavior, is increased (Espinar et al., 2010).

Solar production's intermittent and uncontrollable characteristics cause several other issues, including voltage fluctuations, local power quality, and stability issues. So, predicting solar system output power is necessary for the smooth functioning and the best control of energy fluxes into the solar system, estimating reserves, scheduling the power system, managing congestion, and reducing electricity production costs (Voyant et al., 2017).

Because of the significant increase in solar power generation, forecasting solar yields is becoming increasingly important to avoid significant variations in renewable electricity production. Various systems are being developed to control fluctuations and maintain power quality continuity through forecasting horizons that range from 5 minutes to several days. Moreover, several studies have outlined the reasons to forecast solar radiation for various solar systems. Furthermore, the predicted data's time step will vary depending on the objectives and forecasting horizon (Yang et al., 2020).

In particular, solar systems have the potential to reduce the reliance on carbon-intensive energy sources significantly. Hence, progress over the last 60 years has been made to improve their efficiency. However, due to an anomaly in the output data, current efficiency levels are relatively low. In addition, various abnormalities can occur, preventing solar systems from operating at full capacity. As a result, detecting and repairing these issues is critical to ensure maximum efficiency (Blaga et al., 2019).

In Saudi Arabia, King Abdullah City for Atomic and Renewable Energy (K.A.CARE) collects solar irradiances and the associated meteorological data from 46 ground-based stations across the country to help in the decision-making process related to the country’s projects. The solar irradiance data are collected using different devices that measure incidents on a sensor surface. However, in a hostile climate such as Saudi Arabia, the devices' performance is affected by various elements, such as dust particles accumulation on sensors (Zell et al., 2015) and missing data due to equipment failure, sensor cleaning procedure, or calibration (Vinisha & Sujihelen, 2022). As a result, this can influence data quality severely due to the different anomaly sources.

Since applications depend principally on energy data, such as forecasting energy production or consumption, they necessitated using effective models and accurate input data to give a precise output (prediction). However, with the vast time-series RE data, i.e., solar radiation measurements, it is difficult to assure its correctness. Thus, the most critical task is to confirm that the data collected from different instruments under different weather conditions is accurate. Accurate data will reflect on an accurate model. Thus, the question is how to confirm the data's correctness.

Therefore, this work reviews the literature to understand and analyze the various prediction and anomaly-detection methodologies for big data, i.e., RE or solar data. Furthermore, the work aims to identify the gaps and propose a recommendation for improving the existing techniques to detect data correctness.