Dual-channel encoded bidirectional LSTM for multi-building short-term load forecasting
Vipul Moudgil, Rehan Sadiq, Jagdeep Brar, Kasun Hewage
Abstract
Buildings in a group or cluster orientation account for a rising percentage of electrical consumption causing erratic load patterns and affecting power systems efficiency at grid-level. Short-term load forecasting can provide crucial utility for effective energy transition and optimal energy management with informed decision-making at grid-level. The existing forecasting models predominantly implement deep learning techniques targeting individual buildings. This reduces the model's tendency to account for the inter-building effect while forecasting demand across multiple buildings simultaneously. Also, the existing models often overlook instantaneous peaking patterns of the buildings, leading to sub-optimal learning and predictions. Thus, to address these challenges, this study introduces a novel dual-channel encoded bi-directional long short-term memory (LSTM) load forecasting model. The proposed model is fortified with attention mechanisms, residual connections, and quantile-based metrics that specifically target the erratic peaking patterns across multiple buildings simultaneously. The proposed model is purposely architectured following an encoder-decoder scheme to extract critical local and global temporal patterns and predict electrical demand across multiple buildings simultaneously. The proposed model is tested utilizing the demand data from a Canadian university and the results are compared to the recurrent neural network-based LSTM models. The proposed model improves the average root mean square error ranging from 1.8% to 10.9% and the average mean absolute error ranging from 33.6% to 59% as compared to the existing models. • A novel LSTM network for multi-building short-term load forecasting. • Unified model to predict multi-building load simultaneously. • Use of quantile-based metrics to capture peaking patterns in load data. • Comparison with state-of-the-art LSTM models. • Improves RMSE by 1.8%–10.9% and MAPE by 33.6%–59%.