Enhancing Short-Term Electric Load Forecasting for Households Using Quantile LSTM and Clustering-Based Probabilistic Approach

Abstract

Electricity load forecasting is an essential part of power system planning and operation, and it is crucial to make accurate predictions. The smart grid paradigm and the new energy market necessitate better demand-side management (DSM) and more reliable end-user forecasts to system scale. This paper proposes a time-series clustering-based probabilistic electricity future prediction for short-term load forecasting (STLF), which makes forecasts more accurate and intelligent. The weather and data noise uncertainties are considered, with the load probabilistic interval as the model's output for individual and aggregated household energy consumption. This paper adapts the logarithm of the hyperbolic cosine (logcosh) of the error value as quantile loss and time feature with long short-term memory (LSTM) to bridge the gap. This paper presents a framework for probabilistic electric load forecasting that incorporates clustering-based quantile-LSTM learning to improve the accuracy and robustness of short-term electricity prediction. The proposed model is primarily applied to energy demand information on 15-minute and 1-hour time horizons for day-ahead prediction tasks, which are the key concerns for electricity utilities. Various state-of-the-art regression learning techniques, i.e., quantile regression forest (QRF), quantile regression neural network (QRNN), quantile gradient boosting regression tree (QGBRT), and quantile LSTM (Q-LSTM), were utilized and evaluated. The findings indicate that the proposed cluster-based probabilistic forecasting approach outperforms the existing benchmark models in terms of prediction interval coverage probability (PICP) evaluation metric.