Prediction of Electricity Consumption Profiles Using Potential Polynomials of Degree One and Artificial Neural Networks in Smart Metering Infrastructure

This work analyzes methods and algorithms for predicting the behavior of electricity consumption based on neural networks using data obtained from the Advanced Measurement Infrastructure (AMI) of an educational institution. Also, a contrast between the use of conventional neural networks (ANN), wavelet-based neural networks (WNN) and potential polynomials of degree one (P1P) has been performed. The correlation of each prediction method is analyzed, as well as the behavior of the Mean Square Error (MSE), to finally establish if there is an imbalance in the computational cost through the Big-O analysis and the executing time. The quantitative results of the MSE are below 0.05% for ANN predictions and they use a high computational cost. For P1P, errors around 1.2% are presented, showing as a low computational consumption prediction method but mainly applicable for a short-term analysis. This work is given in response to the need to establish a platform to take advantage of the smart metering structure through the prediction of electricity consumption profile, with the objective of developing a plan for maintenance and management of electricity demand to reduce operating costs from the final consumer to the distribution network operator. For the analysis of projections on the electrical load profile, the statistical characteristics of the consumption are considered to select the prediction algorithms according to the number of days to be projected using data from any of the smart meters, which can be monitored in an electrical network oriented to Smart Grids.