Short-Term Load Forecasting for Smart Grid based on Bidirectional-LSTM Recurrent Neural Network

Abstract

The traditional power grid is evolving into a smart grid, integrating advanced two-way communication technologies and a greater proportion of renewable energy sources, resulting in a more dynamic and flexible network. Accurate load forecasting is crucial for effective operation, planning, and management of the smart grid. Short-term load forecasting (STLF) is particularly challenging due to the high variability and unpredictability in individual consumer behavior, which can impact forecasting accuracy and complicate daily operations and scheduling. Advanced deep learning techniques offer a promising solution to this problem by improving the accuracy of STLF. In this paper, we introduce an ensemble forecasting framework that combines the convolutional neural network (CNN) with a bidirectional long short-term memory (BiLSTM) recurrent neural network with dynamic weight adjustment (DWA). The CNN layers extract features from the data, the DWA layer multiplies the extracted features by their respective dynamic weights before passing them to the BiLSTM model which enhances the forecasting accuracy by capturing both past and future temporal dependencies. We evaluate this framework using a high-resolution real residential smart meter readings dataset and compare its performance against standalone and hybrid models. Our results demonstrate that the BiLSTM-based framework outperforms LSTM-based and traditional approaches in key metrics, including mean absolute percentage error (MAPE) with an improvement of MAPE by 1.99% against the benchmark CNN-LSTM model. This underscores our model's superior accuracy and reliability for STLF, marking a significant advancement over traditional methods. Our model effectively enhances forecasting accuracy in smart grid applications.