Deep Learning-Based Regional Electricity Demand Prediction Using Smart Meter Data in Sri Lanka

Kurzfassung

Precise short-term forecasting of regional and consumer level electricity demand is vital for energy management, especially in countries such as Sri Lanka, where significant energy challenges are faced. However, the increasing availability of smart meters’ data, existing models often lack regional specificity and do not adequately integrate external factors such as weather. This study proposes a robust deep learning framework for the prediction of real-time residential demand in Sri Lanka. A 3.5-year smart meter dataset was utilized, with 15-minute intervals, and combined with weather data. The energy demand being heavily dependent on human behavior, which is difficult to capture but it is significantly influenced by external environmental conditions. The methodology included advanced preprocessing and a generalized feature engineering pipeline, enhancing its adaptability to new regional contexts. Six models based on LSTM and GRU architectures were evaluated using incremental transfer learning strategy. Where models were trained on the Paris dataset and fine-tuned on Sri Lankan data. The analysis revealed different strengths of the model, with LSTM capturing long-term trends, GRU responding to short-term changes, and CNN-hybrids detecting sudden spikes. The GRU with the hybrid BPNN was identified as optimal, achieving the highest R² score of 0.8217 by combining the efficiency of GRU with the capacity of BPNN to model complex nonlinearities. These results were derived using datasets in which anomalies were intentionally retained to preserve real-world operational behaviour. The results confirm the potential of the framework to improve grid operational efficiency and support the development of sustainable energy systems.