Improved Short-Term Load Forecasting Using a Composite Cascaded Neural Network with STL Decomposition

Short-term load forecasting is critical for the operation, control, and planning of power systems. However, achieving high accuracy with existing models remains a significant challenge. This paper proposes a novel approach that combines the STL decomposition algorithm with a composite deep learning network. The STL method is applied to decompose historical load data, extracting trend, seasonal, and residual components. The extracted components exhibit clearer characteristics compared to the original data, and by forecasting these components, the final prediction results are obtained, effectively improving the model’s performance. In addition, by considering the advantages of multiple models, a composite deep learning network is established for short-term load forecasting, which effectively improves prediction accuracy and reduces forecasting errors. Through validation with real-world data, the short-term load forecasting model proposed in this paper demonstrates satisfactory performance and accuracy.