Using a neural network model to determine electricity sales under renewable energy systems penetration consideration

Abstract

Business systems will experience new data-driven models for their performance evaluation in the coming years, especially systems with stochastic characteristics. This development will benefit experts in energy management because more problems will be solved using machine learning algorithms - such as artificial neural networks (ANN). This research develops a machine-learning model for electricity sales using a single hidden layer ANN model. The developed model consists of six input parameters, including the number of renewable energy systems and households. This research used principal component analysis (PCA) algorithm to reduce the inputs to three parameters to improve the model performance. A TOPSIS (Technique for Order of Preference by Similarity to Ideal Solution) method was used to select the most suitable predictive models between SVR (support vector regression) and ANN. Data sets from a community in Lagos, Nigeria, were used to test the developed model performance. This research observed that a SVR model with a linear function performed better than an SVR model with a radial basis function or polynomial kernel. On the other hand, an ANN with 15 neurons outperformed ANN models with fewer nodes. The selected ANN model training and testing mean square errors are 0.00007 and 0.00028, respectively. This research recommends PCA for input parameters selection during electricity sales prediction based on the developed sales model performance.