Short-Term Load Forecasting for Distribution Networks Via Sample Entropy-Guided Secondary Decomposition and Deep Learning

To address the challenges of achieving accurate short-term load forecasting in distribution networks under high load variability and stochastic user behavior, this study proposes a forecasting framework integrating sample entropy-guided secondary decomposition and deep learning. First, the Random Forest algorithm is employed to identify and select the most influential exogenous features affecting load variations. Subsequently, Time-Varying Filtering Empirical Mode Decomposition is applied to perform an initial decomposition of the load series into multiple components. To improve input quality, components with high complexity, as quantified by Sample Entropy, are further decomposed using Singular Spectrum Analysis. Next, the processed sub-sequences and selected features are fed into a bidirectional long short-term memory network, whose hyper-parameters are optimized using an Improved Dung Beetle Optimizer. Finally, the forecasts of all decomposed components are recombined to reconstruct the overall load forecast. Case studies indicate that the proposed approach improves short-term load forecasting accuracy compared with representative benchmark methods for distribution networks.