Improved electric load forecasting using quantile long short-term memory network with dual attention mechanism

Abstract

The robust and accurate load forecasting is necessary to ensure effective power market operations and optimize load dispatch strategies. Deep learning models have recently gained popularity because of their strong ability to learn data patterns. However, conventional deep-learning models still encounter difficulties in precisely predicting complex load patterns. This paper addresses the difficulties of forecasting intricate load patterns, where conventional deep learning models often fail. Therefore, a novel quantile long short-term memory network with dual attention is proposed for hour-ahead short-term load forecasting. By combining dual attention processes with quantile regression-based long short-term memory networks, the proposed framework effectively captures the temporal dependencies of the complex load pattern. The gates recurrent unit and hybridized methodologies of recurrent neural networks are among the baseline techniques against which the proposed method is thoroughly tested using datasets from Panama City and the Islamabad Electric Supply Company. The proposed quantile long short-term memory network with dual attention mechanism has demonstrated notable performance gains with 2.35% and 5.36% reduction in mean absolute percentage error in comparison to the best-performing models, from the set of baseline models, for the Panama and IESCO datasets, respectively. These results demonstrate the proposed method’s effectiveness in providing more improved and accurate forecasts for enhanced grid stability and economic dispatch efficiency. • Improved learning of temporal dependencies with integrated quantile long short-term memory network with dual attention. • Robustness and generalizability analysis with real-world short-term load forecasting application. • State-of-the-art comparative analysis with different performance indicators.

Topics & Concepts