Seasonal variations in the demand and supply of energy in regions around the world make it difficult to accurately forecast energy usage. How should corporations balance the supply and demand for energy, in the presence of the shifting modern restorable energy resources, several weather conditions, and customer usage habits? Yu et al. inspected the distribution of energy usage to develop reliable algorithms for predicting energy consumption.

The authors used the familiar nonparametric Shapiro-Wilk and quantile-quantile plot normality test approaches to investigate the underlining distribution of the meter readings from 283 houses over 200 days. The two test statistics reveal that the energy usage in the mornings, afternoons, and evenings at the houses comes close to a Gaussian distribution.

They examined the unwavering use of a neural network with backward propagation, the traditional support vector machine (SVM) model [1], and a least-squares SVM to forecast energy usage. The historical hourly energy used, humidity, temperature, and wind speed over days were normalized and weighted as features used to train and assess the effectiveness of each machine learning energy prediction model. Not surprisingly, the familiar SVM models produced more reliable forecasts of distributed energy resources, such as wind speeds, over the traditional neural network model. The researchers offer great insights into the transformation of variables that affect energy usage into reliable predictors, and provide statistical measures for identifying malicious energy usage. I strongly encourage all energy providers to read the discerning ideas in this paper, and take advantage of the machine learning models for balancing the periodic supply and demand for energy.