Comparative Analysis of Machine Learning Models for Short-Term Load Forecasting

Load forecasting is challenging owing to the impact of unforeseeable and fuzzy variables such as temperature, humidity, wind speed, and other parameters on load prediction. In this work, algorithms such as Support Vector Machine, Decision Tree, Random Forest, and Linear Regression are used for load forecasting are analyzed for their efficacy. Using the case of each algorithm for a particular load forecasting problem is discussed and general parameter settings for an algorithm to obtain optimal load forecasting are reported. Models are implemented using Jupyter Notebook with Python 3 programming language. Data is used on the Panama Power system from March 2015 to June 2020. This data is further divided into three sizes as small, medium, and large datasets with increasing feature numbers. SVR comes out to be the best algorithm for small datasets. For medium dataset MLR, SVR with holiday feature and RF with holiday performs the best. For large dataset with all features RF perform the best. But the SVR and MLR performs better throughout for all features. Also with temperature feature with large dataset a sudden spike in MAPE from datetime increases by 160.4% for one city temperature, 67.27% for two city temperatures and 25.7% for three city temperatures but when holiday used the same model gives the best results among all model, which shows the importance of holiday feature for SVR in large dataset.