Enhanced Long-Term Prediction Based on 2D Tensorization for Building Energy Consumption

Building energy consumption constitutes a significant contributor to greenhouse gas emissions. Precise forecasting of energy consumption in buildings plays a pivotal role in managing building energy efficiently, thereby aiding in the reduction of greenhouse gas emissions. However, existing prediction algorithms often focus on within-period variations in energy consumption data and overlook between-period variations. This limitation makes it challenging to achieve precise prediction results, especially for long-term forecasts. To address this research gap, this paper proposes an Enhanced Long-Term Prediction based on 2D Tensorization (ELP2T) for building energy consumption. First, a Frequency-Guided 2D Tensorization Network (FG-2TN) is proposed. Energy consumption data, when represented as a one-dimensional time series, faces certain limitations, and FG-2TN is employed to address these limitations. Second, a Progress-Optimized Deep Convolutional Network (PO-DCN) is proposed. It is designed to efficiently extract and learn features from the obtained two-dimensional tensors with fewer parameters and less time. Third, a modular method is proposed to transform one-dimensional time series into two-dimensional tensors using Fast Fourier transform. These tensors are processed and then concatenated into a one-dimensional sequence for output. Ultimately, a comparative analysis was carried out using several traditional forecasting algorithms to highlight the superior performance of the ELP2T model. The obtained average R² score for our proposed method is 0.807, representing an impressive enhancement of 11.06% compared to the most advanced alternatives. This substantial improvement firmly establishes the superiority of our approach. Especially, when considering a prediction length of 720 units, the performance gain of this metric increases to 18.09%, underscoring the pronounced advantage of our method in addressing long-term forecasting scenarios.