Short-term load forecasting for multiple buildings: A length sensitivity-based approach

With the rapid development of large-scale building energy monitoring platforms, it is of great significance to develop precise forecasting methods for buildings on a large scale to achieve better energy system design, system operation, energy management, and renewable energy integration in the grid. Traditionally, using all available historical data to train a data-driven model has been widely employed to ensure prediction performance because more historical information can be learned. However, this strategy may introduce more noise, especially for short-term load forecasting. Thus, this study proposes a novel approach for selectively utilizing building historical data to determine the amount of data that should be used to train the data-driven model. First, the CV(RMSE) curve of each building reflecting the relationship between training data length and forecasting accuracy is obtained using LightGBM. Second, clustering techniques such as k-means are used to identify buildings that are sensitive to the training data length based on CV(RMSE) curves. Finally, the optimal training data length for day-ahead forecasting is estimated for each building. The case study shows that approximately 20% of buildings in the Building Data Genome are labeled as length-sensitive buildings, and adopting appropriate training data lengths can reduce the prediction error of these buildings by up to 15%.