Forecasting highly fluctuating electricity load using machine learning models based on multimillion observations. (August 2022)

• Record Type: Journal Article
• Authors: Abdallah, Mohamed; Abu Talib, Manar; Hosny, Mariam; Abu Waraga, Omnia; Nasir, Qassim; Arshad, Muhammad Arbab
• Is Part Of: Advanced engineering informatics. Volume 53(2022)
• Publication Date: 2022-08
• Subjects: Machine learning -- Cross-validation -- Electricity consumption -- Ensemble models -- Artificial neural network; Computer-aided engineering -- Periodicals; Engineering -- Data processing -- Periodicals; 620.00285
• DOI: 10.1016/j.aei.2022.101707
• Languages: English
• ISSNs: 1474-0346

Abstract

Highlights: 50 million observations were used to model monthly electricity consumption. Effect of external features on electricity consumption was modeled and investigated. Randomly split and temporally ordered datasets were used to examine the AI models. Random forests achieved the highest accuracy when trained using both datasets. Rolling and moving cross-validation were implemented to analyze the models. Neural network was the most stable model with minimal variation in predicted values. Abstract: Dubai is an emerging metropolis with unique features, such as an extremely high expatriate ratio and turnover rate. These features, along with the extreme climatic conditions and dependency on energy-intensive water desalination, cause the electricity load to be high per capita and highly variable. Therefore, producing accurate electricity consumption predictions is critical to prevent blackouts or power loss, particularly for fast-growing cities. This study predicts the monthly electricity consumption in Dubai using various linear regression variants, support vector machines, decision tree models, ensemble models, and neural networks. The dataset consists of different demographic and climatic features as well as more than 50 million electricity consumption observations aggregated in monthly community-wide records. By analyzing the correlation between the various features and electricity consumption, a set of 18 factors was determined, with the number of expat sites, the total number of sites, and population identified as the most influential variables. The models examined were trained on the selected features, and their prediction performance was evaluated using various evaluation metrics. The results showed that the ensemble models were best in predicting consumption. Random forest outperformed the rest when trained using a temporally ordered dataset with 0.0437 and 0.0226 as root mean square error and mean absolute error values, respectively. In addition, the employment of a randomly split dataset resulted in better performance. Ensemble models were found to predict electricity consumption accurately when trained on 60% of the historical consumption records. The generalization and reliability of the top models were further tested by varying the training data volumes and applying rolling and moving cross-validation. Training the models on a fixed-size moving window of historical records leads to better performance and more reliable prediction.