Accurate multi-horizon load forecasting is essential for the stability and efficiency of smart grid operations, particularly in residential environments where electricity consumption patterns are shaped by both long-term trends and short-term fluctuations. Transformer-based models such as Autoformer have advanced forecasting accuracy by leveraging frequency-domain attention to capture periodic behavior. However, they often struggle with rapidly changing, localized patterns prevalent in real-world data. To address this challenge, we propose CLM-Former, a novel hybrid deep learning architecture that integrates time series decomposition, an autocorrelation-based attention mechanism, and a tailored subnetwork, CLM-subNet, which combines convolutional and recurrent layers. This design enables the model to effectively capture both seasonal dependencies and high-resolution variations in electricity usage, thereby enhancing its performance across multiple forecasting horizons. Comprehensive evaluations on real-world smart meter data demonstrate the robustness and adaptability of CLM-Former against a range of Transformer-based and deep learning baselines. By effectively modeling both long-term periodic trends and short-term dynamics, CLM-Former emerges as a promising tool for residential energy forecasting. Its robust performance offers valuable implications for demand response, distributed scheduling, and the future management of smart grids.

Keywords