LightWaveDual: A Knowledge-Guided Linear Dual-Branch Network for Building Natural Gas Consumption Prediction

Accurate prediction of natural gas consumption in buildings is crucial for improving energy efficiency and achieving carbon neutrality goals. However, building energy consumption prediction faces significant challenges due to complex energy behaviors, multi-periodic patterns in data, and the limitations of existing deep learning methods which often require high computational overhead. This paper proposes LightWaveDual, a lightweight dual-branch linear prediction framework that reformulates time series prediction as a multi-scale signal processing task. Our key innovations include: 1) AdaWave, a wavelet-like linear convolution module for frequency component separation; 2) CyclicPredictor, a period-aware architecture enhancing multi-scale temporal features; and 3) SchedulerAwareLoss, an energy scheduling-oriented loss function combining modified Huber loss and volatility loss. Experiments on the UNICON dataset show that our model outperforms existing methods across multiple prediction horizons (1–7 days), achieving a Mean Absolute Error of 3.1745 for one-day prediction, a 2.4% improvement over baselines. Ablation studies validate each component's effectiveness, particularly in long-term prediction accuracy. This research provides an efficient solution for building energy consumption prediction with practical value for energy conservation.