Industrial Power Load Forecasting for Grid Operation Using a CNN-Transformer-BiLSTM Model

Abstract

In the time of AI era, Industrial power load gradually rising due to the rapid expansion of the chip manufacturing facilities. So that accurate forecasting of industrial power load is important to achieve efficient grid planning and overall energy management. But, due to the nonlinear, volatile and multi scale nature of industrial power load data, the conventional statistical model face challenges in forecasting efficiently. To address these challenges, a novel hybrid deep learning model, CNN-Transformer-BiLSTM has been proposed that integrates the feature extraction capacity of convolutional neural networks (CNN), the long-range dependency modeling of the transformer architecture and the sequential learning strength of bidirectional long, short-term memory (BiLSTM) networks. The CNN layers efficiently capture the local temporal patterns and feature correlations within the load data sets, Transformer layers employs self-attention mechanisms to model complex long-term dependencies and contextual relationships. The BiLSTM layer further enhances temporal representation by learning bidirectional dependencies, thus improving the overall prediction accuracy. Historical monthly industrial electricity load data from the U.S. Energy Information Administration (EIA) spanning over two decades are used to train and evaluate the model. The proposed model output has been compared with other standalone and hybrid deep learning models. The proposed CNN-Transformer-BiLSTM achieves superior forecasting accuracy with Mean Absolute Percentage Error (MAPE) of 1.23%, Root Mean Square Error (RMSE) of 1,276 MWh and Mean Absolute Error (MAE) of 1,040 MWh.