Forecasting photovoltaic power in high-latitude regions via support vector machine optimized by meta-heuristics

Machine-learning techniques are widely used across many disciplines, including electricity generation forecasting. In this study, the Support Vector Machine (SVM) based models, one of the machine learning techniques, were developed for daily PV power forecasting. To improve model performance, models were tuned with four metaheuristic optimizers, including the Artificial Bee Colony (ABC), Grey Wolf Optimizer (GWO), Genetic Algorithm (GA), and Particle Swarm Optimization (PSO). Daily PV power and temperature data from 2020 to 2023 were obtained for the Stavanger, Oslo, and Kristiansand regions which located in southern Norway. One of the innovative aspects of this study is the investigation of the performance of SVM (Support Vector Machine) combined with various optimization methods across four alternative input configurations. To examine the different feature combinations, four different input configurations were created through the Minimum-Redundancy Maximum-Relevancy (MRMR) method. The analysis results obtained with SVM were further enhanced using all optimization techniques. Among those, the SVM-PSO-M04 (r = 0.7707, NSE = 0.5748, KGE = 0.7092, PI = 0.2964 and RMSE = 0.6513) method produced the most effective results (improving the correlation coefficient (r) to 0.7707 (approximately a 19% increase over the untuned SVM)) among the tested hybrid configurations obtained in our experiments. Moreover, coupling temperature data alongside PV power as model input also tends to improve forecasting skill. Results of this study provide a case-study benchmark for researchers, institutions, and other stakeholders engaged in renewable energy planning and management in high-latitude regions.