Applicability of data-driven methods in modeling electricity demand-climate nexus: A tale of Singapore and Hong Kong

Abstract

Predicting electricity demand at city level is vital for stakeholders. There are various data-driven methods in electricity consumption prediction, but their applicability to forecast monthly electricity demand has yet to be systematically explored. This study examines several widely used data-driven methods, namely multiple linear regression (MLR), machine learning (ML) method including support vector machine (SVM) and random forest (RF), and deep learning (DL) method including long short-term memory network (LSTM) and LSTM-gated recurrent unit (GRU), for long-term energy prediction using climatic and historical electricity datasets in Singapore (2005–2019) and Hong Kong (1975–2019). In Singapore, ML outperforms other methods in terms of statistical criteria and time series plots, and SVM provides the best accuracy with mean absolute percentile error (MAPE) of 2.55 %, while MLR exhibits the worst accuracy with MAPE of 3.1 %. In Hong Kong, DL surpasses ML in terms of statistical criteria, time series stability, and generalization ability. MLR achieves the best overall prediction accuracy with R2 up to 0.95 but shows poor ability for predicting peak and low electricity consumption, while LSTM exhibits the smallest bias in these months. RF shows strong overfitting issues in both cities. Overall, SVM and LSTM are recommended for small and large datasets, respectively.