<![CDATA[Short-Term Load Forecasting (STLF) is essential for maintaining grid stability and optimizing operational efficiency in modern energy systems. While traditional Convolutional Neural Networks (CNNs) can extract local temporal features, they often struggle with capturing long-term dependencies and demand high computational resources. This study proposes a novel application of the Ghost Convolutional Neural Network (GhostCNN)—initially designed for image processing—to time-series electricity load forecasting. GhostCNN significantly reduces model complexity while preserving forecasting accuracy by generating redundant temporal features through lightweight linear operations. The model is trained and evaluated on a real-world electricity load dataset from Ho Chi Minh City, containing 13,440 hourly observations (~1.5 years). A comprehensive hyperparameter tuning strategy is applied, covering kernel size, Ghost ratio, sequence length, batch size, and learning rate. The model's performance is benchmarked against MLP, CNN, and LSTM architectures. GhostCNN achieves the lowest Mean Absolute Percentage Error (MAPE) of 1.15%, outperforming CNN (1.27%), MLP (1.67%), and LSTM (7.3%). Furthermore, GhostCNN reduces inference time by approximately 40% and decreases parameter count by ~45% compared to standard CNNs, affirming its suitability for real-time smart grid deployment. These results demonstrate that GhostCNN provides a robust, scalable, and efficient solution for accurate short-term electricity load forecasting in dynamic and resource-constrained environments.]]>
Copyrights © 2025
