Super Hybrid Deep Learning Framework: Transformer-LSTM-CNN-DNN for Accurate Load Forecasting

The short-term demand forecasting of electric energy allows today’s power grid operations to run without interruptions. The improvements of Machine Learning and Deep Learning together with higher computational performance, have considerably improved the precision of the forecasts. In Bangladesh, where effective resource management can be a crucial factor, power load forecasting plays a vital role in energy security. In this research, a hybrid feature generation model using a combination of Transformer, LSTM and CNN which was then fed into a DNN model, was built for a robust predictive performance. The study was conducted on historical daily load data, spanning the period from January 2014 to April 2023, covering all eight divisions of Bangladesh, which was also subjected to a series of data cleaning, normalization and partitioning into training and testing subsets. Validation of the model was done for all the divisions, with Sylhet and Rajshahi divisions yielding the best R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.993 and 0.989, respectively. The Transformer, LSTM and CNN layers captured both long-term dependencies and short-term variations, while the prediction accuracy of the DNN model provided a robust approach to load forecasting. This super-hybrid approach accentuates the benefits of flexible modern deep learning models to the electricity distribution system, especially for countries like Bangladesh, where efficient utilization of energy resources is crucial due to limited resource and growing demand.