Monthly Electricity Consumption Prediction: Integrating Artificial Neural Networks and Calculated Attributes

Abstract

Electricity consumption is increasing on a daily basis, and consequently, the need for its control, potential reducing or at least predicting, is growing. The aim of this research is to predict the electricity consumption based on consumer attributes, using a dataset with a poor list of useful attributes as a starting point. Even though the electricity distribution company from which the data were obtained records data on electricity consumption precisely, the obtained data did not provide enough information to ensure a satisfactory level of the estimation precision. That is why, for the purpose of this research, the initial dataset was subjected to the extensive treatment in the preprocessing phase and updated with a lot of additional, collected and calculated attributes. Subsequently, the neural network model that predicts electricity consumption on a monthly basis was proposed. Basically, two models were created, with several variations in the number of neurons in the hidden layers, but with the identical structure of input and output layers. The proposed models were tested on a very complex dataset, obtained by updating the initial one, and comprising all the measuring points and all types of consumers in the area of the City of Užice, recorded during a period of 56 months. The results show that the proposed metodology of updating a dataset with additionaly collected and calculated inputs, together with the proposed neural network model, ensures a very low prediction error, i.e., ≈5%. This could make electricity consumption control and reduction, but also electricity production planning possible.