Tracing the sources of online prediction errors in building energy consumption: A bias-variance and time-frequency domain based approach

The online prediction of building energy consumption is crucial for energy system management, operation, and real-time control decision. However, the complexity and uncertainty of building operation conditions often leads to prediction failures whose sources are difficult to trace, limiting practical applicability. This study proposes a novel method to identify error sources in online prediction by combining bias-variance decomposition with time-frequency characteristic analysis. The approach classifies error types and examines the underlying data factors in deep model, thereby enabling error traceability. This paper utilizes energy consumption data from four representative public buildings in China to conduct short-term online predictions of seasonal energy consumption. The results demonstrate that the proposed method can proficiently identify three prevalent types of prediction failure, namely strong bias, high variance, and data misalignment. Specifically, prediction failure becomes more likely when the standard discrete coefficient of training data and the weighted Euclidean distance of feature variables surpass certain thresholds, while high daily cycle intensity mitigates the risks. These findings highlight the effectiveness of the proposed error source identification method as a practical tool for resolving the challenges of unidentified error sources in online prediction. Overall, this study establishes a practical and interpretable methodology for diagnosing error sources in online building energy prediction, offering new insights for adaptive modeling and future digital twin applications.