Near real-time machine learning framework in distribution networks with low-carbon technologies using smart meter data

The widespread adoption of low-carbon technologies, such as photovoltaics, electric vehicles, heat pumps, and energy storage units introduces challenges to distribution network congestion and power quality, particularly raising concerns about voltage stability. Enhanced voltage visibility in low-voltage networks is increasingly vital for active grid management, making efficient voltage forecasting tools essential. This study introduces a novel data-driven approach for forecasting node voltages in low-voltage networks with high penetration of low-carbon technologies. Using time series of power measurements from smart meter data, the study integrates an Extreme Learning Machine with the Single Candidate Optimizer to enhance computational efficiency and forecasting accuracy. The model is validated using smart meter datasets from two different low-voltage networks with low-carbon technologies and is compared with several established machine learning models. The results demonstrate that the optimization algorithm significantly improves the tuning of model parameters, achieving up to a 17-fold reduction in computation time compared to the fastest metaheuristic methods implemented. The proposed model demonstrated superior accuracy, with an average voltage deviation of 0.56%. Although the computation time per node achieved is not yet suitable for real time applications, the study shows that the optimization method significantly improves the performance of the forecasting tool. • A novel data-driven model optimizes voltage forecasting in LV networks with LCTs. • The ELM model parameters are enhanced using the SCO metaheuristic method. • The proposed model significantly reduces computation time up to 17-fold. • Superior voltage forecasting accuracy with a deviation of just 0.56% is achieved. • The study uses real-world smart meter data collected from an LV network in Ireland.