Modeling and Empirical Evaluation of Machine Learning Based Load Forecasting Models for Pakistan

Abstract

Electric load forecasting (LF) deals with predicting futuristic energy demand of consumers. It is the foremost and important step of energy distribution and generation planning. Machine learning based statistical and artificial intelligence techniques are widely used for LF. Among these, artificial neural networks (ANN) and support vector machines (SVM) emerge as competitive modeling approaches for LF. To further improve the performance of these models, optimization techniques are being used to formulate hybrid LF models. Availability of modern approaches motivated authors to solve the issues with power planning in Pakistan. Hence, we contribute towards proposing machine learning based accurate model of LF on Pakistan power system data set. Several forecasting models are formed using hybrid optimization and model development techniques, which are ranked against their forecasting accuracy and performance. SVM based models performed well and achieved 98.91% accuracy of forecasts. On the other hand, ANN based models showed comparable performance achieving 98.34% accuracy with added ability to avoid over-fitting, and efficiency with improved results.