Machine learning based photovoltaic output power forecasting

The installation capacity of photovoltaic (PV) systems is gradually increasing in order to prevent reasons such as limited fossil fuel resources, harmful effects on the environment and the bad course in the global energy field The dependence of PV systems on weather conditions causes instability, voltage, frequency fluctuations and interruptions in PV power outputs. This situation complicates the integration of PV energy into the grids. Therefore, short-term forecasting of PV power output is crucial to overcome the challenges. The aim of this study is to use the Robust Arranged Random Vector Function Interconnect (GD-RVFL) network model, which learns faster and performs with high accuracy, overcoming the excessive learning and slow learning disadvantages that are common in machine learning models in the literature, to predict the short-term PV output power and in this context by comparing the proposed model with Bayesian Ridge Regressor (BRR), Linear Regressor (LR), Gaussian Process Regressor (GPR), Support Vector Machine (SVM), Extreme Learning Machine (ELM), Artificial Neural Network (ANN), Gradient Boosting Regressor (GBR), Random Forest Regressor (RFR), Lasso Regressor (LAR) and Ridge Regressor (RR) methods, which are 10 different machine learning methods to evaluate the performance of the models. In this study, MATLAB program was used to create and implement machine learning models. The data set was collected from the PV power plant and the solar power station located on the Dicle University campus in Diyarbakır. The dataset has 10-minute periods between May 2019 and April 2022. First, solar radiation, sunshine duration, wind speed, ambient temperature, panel surface temperature, cloudiness, DC current, voltage and relative humidity parameters were divided into four classes according to weather type levels using Conical Correlation Analysis (KKA) model. Next, the RR-RVFL model estimated the short-term (10 minutes) PV output power using these four classes of air type as input, and the results were compared with 10 different machine learning models. As a result of this comparison, it was seen that the efficiency of RR-RVFL significantly outperformed the other 10 machine learning models. Also, on days when meteorological conditions in the region were relatively stable, most of the models showed relatively high accuracy. On days when weather conditions varied, RR-RVFL gave better results. Mean Absolute Percentage Error (OMYH) values of the days with constant meteorological conditions were 1.68, 1.72 and 1.82 respectively. As a result, the selected features and method significantly improved the prediction accuracy.