Multivariate short-term load forecasting based on hybrid deep learning approach

Short-term Load Forecasting (STLF) is crucial in planning the safe, reliable and efficient operation of today’s power systems. However, the variability arising from the source load pattern due to increasingly dynamic external influence creates considerable hurdles in precise forecasting. This study introduces an innovative hybrid forecasting model that integrates three state-of-the-art deep learning architectures: Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), and Residual Networks (ResNet) are some of the subtypes of Artificial Neural Networks that have been implemented in deep learning networks. To achieve the goal, the temporal learning ability of LSTM is combined with the feature extraction ability of CNN and the efficient learning ability of ResNet, to capture temporal and spatial features in load data. A preprocessing plan is defined to enhance the quality of data to deal with the anomalies results from the outlier, zero value imputation and normalisation stages. The proposed hybrid model is then analysed and tested with actual load dataset through cross-validation to see performance degradation The proposed hybrid model also cross validated with benchmark standalone LSTM, CNN, ResNet models and traditional machine learning technique. In case of hyperparameters, Bayesian optimization is involved in the last stage to fine tune the results of the model. The analysis shows that the suggested hybrid model yields notably higher levels of prediction accuracy and resilience than all other baseline scenarios. This work demonstrates how proper design of hybrid deep learning architectures can eliminate the difficult of STLF and thus ensure better operation of power systems.