Syner-Dandelion Optimization Integrated Temporal Recurrent Transformer Network for Accurate Energy Load Forecasting in IoT-Smart Grids

Accurate energy forecasting is central to providing optimal functioning of Internet of Things (IoT)-smart grids in predicting the correct amounts of consumption, facilitating optimal resource allocation, and providing grid stability. Increasing the integration of renewable sources and decentralized generation, real-time energy demand forecasting is more important than ever to providing precise sustainable and resilient energy supply. However, classical forecasting methods are afflicted with three widespread limitations: 1) inadequate management of timely IoT data; 2) low predictability for intricate energy system patterns; and 3) weak modeling of spatiotemporal correlations. To meet these urgent needs, this article proposes a new deep learning-based energy load forecasting approach blending IoT data harmoniously with a hybrid structure. The proposed temporal recurrent transformer network (TeRT-Net) integrates the sequential learning ability of recurrent neural network and global attention of Transformers in capturing long-distance dependence and salient time points of energy consumption behavior. Syner-Dandelion Optimization (SyD-Opt) is also included in the approach to optimize model training with the aim of escaping local minima and enhancing convergence. Experimental results verify that the method under consideration is well capable of enhancing forecasting performance, particularly during peak load hours and variability caused by uncertainty in renewable energy. The research contributes much to researchers and practitioners as it introduces a low-cost scalable method for real-time energy management and intelligent decision-making applications in a smart grid. The novel SyD-Opt combined TeRT-Net model obtained remarkable enhancements in energy load forecasting accuracy, evidenced by a root mean square error (RMSE) of 0.0043, mean absolute error (MAE) of 0.0031, and mean absolute percentage error (MAPE) of 1.08% on the individual household electric power consumption (IHPEC) dataset. These outcomes surpass baseline models like Bi-LSTM (RMSE: 0.0098, MAPE: 2.11%) and CNN-LSTM (RMSE: 0.0145, MAPE: 2.75%), which attest to the fact that the model predicts more intricate temporal patterns and minimizes prediction errors for smart grid usage.