Rule-based Classification and Outlier Replacement for Daily Electricity Load Forecasting

This paper presents rule-based classification and outlier replacement for data arrangement and preprocess for daily electricity load forecasting. The historical load data from 2019 to 2021 are provided by the Electricity Generating Authority of Thailand (EGAT) and there are 48 periods in each day. The training data from 2019 to 2020 and the test data from 1st January 2021 to 31st December 2021 are applied. In both datasets, the independent variables include historical load (one day before, load one week before, and two weeks before). The rule-based classification is presented where the data is classified manually to group similar load patterns and cleaned by replacing holidays, bridging holidays, and outliers. The mean absolute percentage is used to evaluate the performance of the proposed model. We use the MLR, SVR, XGBoost, and NN models to apply this approach. Furthermore, we propose the ensemble approach that combines the strengths of these powerful models. The results show that the proposed MLR model is more effective in forecasting performance than others.