DEVELOPMENT OF A PREDICTION SYSTEM FOR HOUSEHOLD ELECTRICITY CONSUMPTION IN DHAKA CITY

Abstract

In today's modern world, electricity plays a crucial role in everyday life, serving as an essential resource for various activities. With uninterrupted access to electricity being a top priority for nations worldwide, accurately forecasting electricity consumption has become increasingly important. Such forecasts are instrumental in enhancing the reliability and efficiency of power supply infrastructure. A review of the relevant studies shows that, currently there is no application with which the electricity consumption of Dhaka city can be predicted. In order to fill in the need for such an application, the primary objective of this research was - conducting performance analysis among state of the art machine learning algorithms in terms of predicting power. In order to accomplish this objective, consumption data of a group of consumers from Dhaka city (specifically in Kafrul, Kallyanpur, Pallabi, Monipur and Rupnagar area) was collected for a period of three years. Then the data was used to train machine learning models for predicting power consumption with Artificial Neural Network, Support Vector Machine, XGBoost, LightGBM and CatBoost. In these models, features like monthly consumption (kilowatt-hour) for two years, sanctioned load, meter type, etc. had been selected to predict the electricity consumption of next one year. Grid search technique had been deployed to find the appropriate hyper-parameter values for each algorithm. Afterwards, the performance of each of the proposed predictive models was measured to determine their effectiveness by calculating MAE, RMSE, R2 values for each algorithm. Experimental results present that Artificial Neural Network (ANN) and Support Vector Machine (SVM) shows more accurate result compared to the other algorithms as per MAE, RMSE, and R squared method. For example, the MAE for ANN and SVM were 49.19 and 51.08 respectively, where the MAE for XGBoost, LightGBM and CatBoost were 127.22, 165.62 and 70.81 respectively. However, in terms of time required to train the model, LightGBM and CatBoost stands out among all these algorithms. The secondary objective of this research was to develop a web application using the best performing model to predict the electricity consumption of the consumers of Dhaka city. To achieve this objective, a web application had been developed using the model trained with Artificial Neural Network (ANN) where the users can put the electricity consumption of any consumer during two years and predict the electricity consumption for the next one year. Output of the experiment shows that adopting the methods may greatly benefit power generation, transmission and distribution entities as well as the government in forecasting energy needs and plan accordingly. The methods used in this work can also be utilized to reduce system loss and prevent equipment damage to a great extent.