Hybrid KNN-SVM machine learning approach for solar power forecasting

Abstract

Predictions about solar power will have a significant impact on large-scale renewable energy plants. Photovoltaic (PV) power generation forecasting is particularly sensitive to measuring the uncertainty in weather conditions. Although several conventional techniques like long short-term memory (LSTM), support vector machine (SVM), etc. are available, but due to some restrictions, their application is limited. To enhance the precision of forecasting solar power from solar farms, a hybrid machine learning model that includes blends of the K-Nearest Neighbor (KNN) machine learning technique with the SVM to increase reliability for power system operators is proposed in this investigation. The conventional LSTM technique is also implemented to compare the performance of the proposed hybrid technique. The suggested hybrid model is improved by the use of structural diversity and data diversity in KNN and SVM, respectively. For the solar power predictions, the suggested method was tested on the Jodhpur real-time series dataset obtained from the data centers of weather stations using Meteonorm. The data set includes metrics such as Hourly Average Temperature (HAT), Hourly Total Sunlight Duration (HTSD), Hourly Total Global Solar Radiation (HTGSR), and Hourly Total Photovoltaic Energy Generation (HTPEG). The collated data has been segmented into training data, validation data, and testing data. Furthermore, the proposed technique performed better when evaluated on the three performance indices, viz., accuracy, sensitivity, and specificity. Compared with the conventional LSTM technique, the hybrid technique improved the prediction with 98 % accuracy.