Transfer-learning enabled adaptive framework for load forecasting under concept-drift challenges in smart-grids across different-generation-modalities

Abstract

The Smart Grids (SGs) have significantly improved the load demand with the help of different generation modalities (DGMs), that are supporting the energy demand. Equally, it has increased the complexity in the electrical systems due to multi data flow. This continuous flow of data from multiple sources influences concept drift (CD) in the data streams. This CD causes challenges for forecasting systems to efficiently interpret the electrical demand and deteriorate the forecasting performance of existing models. This study presents a transfer learning enabled adaptive long short-term memory (TLA-LSTM) framework for electrical load forecasting for DGMs in Malaysia. The study utilizes extensive datasets from four distinct modalities i.e. Coal, Gas, Hydro, and Solar (MYC, MYG, MYH, and MYS) and evaluates the performance of TLA-LSTM under three different case studies of CD: Before Concept Drift (B CD ), Concept Expansion (C EXP ), and Concept Evolution (C EVO ). The TLA-LSTM demonstrated superior forecasting accuracy, consistently achieving higher R 2 score and lowest normalized root mean square (NRMSE) and mean absolute percentage error (MAPE) compared to traditional models like random forest (RF), support vector machines (SVM), standard LSTM, and adaptive ensemble LSTM (AE-LSTM). The results of TLA-LSTM show an improvement of approximately 7–15 % under different CD scenarios. These results highlight TLA-LSTM advanced architecture, which effectively captures both temporal dependencies and non-linear relationships, leading to significant improvements in forecasting accuracy. This paper examines these trade-offs, providing valuable insights for researchers and practitioners in the field of electrical load forecasting under CD.