Hybrid support vector regression with clustering approach for Thailand short-term electricity load forecasting

Accurate forecasting of electricity load demand is crucial for effective management in the energy sector. However, due to the high level of uncertainty involved, achieving precise electricity load demand forecasts presents a significant challenge. Several factors, such as historical electricity demand, seasonal effects, day of the week, and holidays, influence the electricity. A combination clustering and an optimized support vector regression algorithm is proposed to improve short-term electricity load forecasting.The electricity load demand data is collected from the Electricity Generating Authority of Thailand (EGAT) and divided into two sets: a training set consisting of electricity load data from April 2019 to March 2020, and a testing set consisting of data from April 2020 to March 2021. The training dataset is used to apply K-Means clustering, which groups the electricity load demand based on their load patterns. For the testing set, a K-Nearest Neighbors classifier is employed to predict the cluster group label based on various features, including the day of the week, the month of the year, the holiday, and the bridging holiday.After clustering the training and testing datasets, the electricity load demand is predicted using Support Vector Regression with Bayesian optimization (SVR-BO). The proposed model is then compared with deep learning , manual classification based on days, and the load demand is forecasted using SVR-BO. The Mean Absolute Percentage Error (MAPE) metric measures forecasting accuracy.The proposed model in this study achieves remarkable performance in forecasting holiday load demand, except for the months of January and December. Moreover, the model outperforms manual classification in predicting weekday and weekend load demand. The main contributions of this research are integrating clustering and forecasting methods, applying K-Means clustering only on the training data, and using the K-Nearest Neighbors classifier to assign the correct cluster label based on relevant external factors for the testing data.