Highly fluctuating short-term load forecasting based on improved secondary decomposition and optimized VMD

Short Term Load Forecasting (STLF) is a critical task in the power sector, enabling efficient resource allocation and grid management. However, the volatile and complex nature of short-term load series pose significant challenges to forecasting models. Traditional decomposition-prediction models are bottlenecked in that they often lack complexity-based clustering for efficiency and optimization of decomposition for optimal secondary decomposition. In this paper, we summarize the framework of the decomposition-prediction models, and propose the hybrid model to address these limitations. We propose a Sample Entropy-based hierarchical clustering method to cluster components according to complexity and improve the efficiency of secondary decomposition. Additionally, we propose the center frequency method to efficiently optimize the K parameter of VMD, ultimately achieving the optimal decomposition. In summary, firstly, to help minimize the difficulty of prediction, the load series is decomposed twice using Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and Optimized Variational Mode Decomposition (OVMD). Then, two separate Long Short-Term Memory (LSTM) frameworks are built to predict the components obtained from the two decompositions, thus leveraging the advantages of the previous basic framework. Finally, by superimposing the prediction results, we obtain the output of the proposed model. The Belgian power load dataset is divided into four groups by season for comparison experiments. The results reveal that our model outperforms the benchmark models, with the best average coefficient of determination and mean absolute error of 0.996 and 53.69. Additionally, the limitations of sample entropy in secondary decomposition were revealed through our findings. These insights emphasize the promising contribution that our study brings in enhancing the decomposition-prediction model.