Machine learning-based energy consumption forecastingfor stores in a shopping center - A case study

Abstract

Short-term power load forecasting plays a crucial role in the management of power systems. It serves as the foundation for network structure planning, electricity trading, and load scheduling. The accuracy of power load forecasting directly impacts the security, stability, and economic efficiency of the power system. Accurate short-term load forecasting (STLF) is particularly important for the planning and commitment of power plant units. It helps mitigate the uncertainty introduced by the intermittent generation of renewable sources. In today's world, economic growth heavily relies on the availability of electric energy, as societies, industries, and economies are highly dependent on its continuous use. Therefore, having a reliable, affordable, and uninterrupted energy source is of utmost economic significance. Electric load forecasting plays a vital role in ensuring that utilities can meet the energy needs of consumers. To accomplish this, a team of trained professionals is required to carry out this specialized task. Load forecasting encompasses predicting the future load on a specific system over a defined period. These predictions can range from fractions of an hour for operational purposes to up to 20 years into the future for planning purposes. The objective of this master's thesis is to use accurate machine learning models for shortterm electric load demand forecasting in the context of a store in a shopping center. The load data used in this study was collected over a mid-term duration from July to August 2022, specifically during the summer season. Throughout this period, the air conditioning (AC) unit operated daily to maintain indoor temperatures during the center's working hours. The datasets were collected from a store in a shopping center in Istanbul, Turkey, preprocessed, cleaned and prepared to analysis. Furthermore, I utilize six powerful forecasting techniques: artificial neural network (ANN), k-nearest neighbors (KNN Regression), long short-term memory (RNN-LSTM), extreme gradient boosting (XGBoost), and autoregressive integrated moving average (ARIMA) to overcome nonlinear problems in short-term load forecasting (STLF). The models were evaluated using an everyday training dataset consisting of 80 % training data and 20 percent testing data. The results indicates that the LSTM model outperforms in a high accuracy in comparison to other models with accuracy of 95 % in training and 94% in validating. The study shows that the LSTM model is highly effective in load forecasting. The optimization of model parameters (hyperparameter tuning) through trial and error greatly improved accuracy. These findings contribute to load forecasting by highlighting the superior performance of the LSTM model and the importance of parameter selection for accurate predictions.