MoADL-TLSTM: Multiobjective Automated Deep Learning-Based Transformer-LSTM for Load Forecasting

Accurate load forecasting serves as the critical factor for optimizing energy allocation and ensuring economic operation in cyber-physical power systems (CPPSs). Deep learning (DL) models have emerged as pivotal tools for capturing complex temporal patterns for enhancing the performance of load forecasting. However, the existing DL-based load forecasting models contain some shortcomings such as predominant reliance on single DL model neglecting hybrid integration potentials, manual design requiring substantial domain expertise and high model complexity lacking lightweight. To address these critical challenges, a novel multiobjective automated DL (MoADL)-based transformer-long short-term memory (TLSTM) load forecasting method is proposed to automatically design a hybrid and lightweight model by keeping the balance between model performance and model complexity, called MoADL-TLSTM. In MoADL-TLSTM, we combine Transformer's global dependency modeling with LSTM's sequential processing capabilities, enabling feature extraction from load sequences. Then, we develop the nondominated sorting genetic algorithm II (NSGA-II)-based evolutionary mechanism with thevariable-length encoding scheme and specifically designed crossover and mutation operations to represent and evolve the neural architectures (NAs) and hyper-parameters (HPs) of the TLSTM. The experiments use 12 real-world load datasets from the Australian Energy Market Operator on three Australian regions over four seasonal periods. Experimental results demonstrate the superiority of the proposed MoADL-TLSTM method to other state-of-the-art methods included manually designed single DL models, hybrid DL models, and single-objective automated DL (ADL)-based models in terms of model forecasting performance and model complexity.