High Precision Modeling of Power Demand for Electricity Distribution Companies with frequent load shedding

Abstract Globally, electricity forecast plays a critical role in ensuring the stability of the national grid. However, developing counties are challenged with a lack of data infrastructure for continuous data capture, storage, and process that ensure quality data are available for short-term power forecasting, a critical requirement for maintaining grid stability but at the same time utilizing loading scheduling for unbalanced power distribution and grid stability. This led to a huge gap and significant paucity in existing historical power demand. This research provides a mythological framework for modeling and missing value imputation of mean, median, and interpolation while combining meteorological data set from NASA and Power demand datasets from Distribution companies, an approach for improved short-term forecasting accuracy. This research considered Linear machine learning algorithms which include Linear regression, Lasso Linear Regression, Elastic Linear Regression, and nonlinear machine learning algorithms which are Support Vector Machine, Random Forest, Decision Tree, and K- Nearest Neighbor and deep learning algorithms like Artificial Neural Networks and Long Short Term Memory. Among these algorithms, the Deep learning method outperformed all of them with a performance accuracy of above 82.2% for ANN and with 91.2% performance which is prone to overfitting. Hyperparameter tuning with a batch size of 50 and an epoch of 5 gives the best performance for the deep learning methods.