Representation Vector-Based Time Series Similarity Analysis Using Unsupervised Contrastive Learning Time Series Encoder

Data scarcity and high development costs pose significant challenges to building-specific energy demand forecasting models. To address these issues, this study introduces a time series similarity assessment method that utilizes TS2Vec, an unsupervised learning-based encoder for extracting time series representation vectors. The efficacy of this approach is demonstrated using anonymized datasets of building electricity usage from Cambridge, UK. The proposed methodology stands out for its ability to identify high-similarity data segments by flexibly adjusting the evaluation time window used for extracting representation vectors, outperforming traditional average similarity assessments. Principal component analysis was employed for dimensionality reduction and visualization, alongside a moving window cosine similarity approach to enhance the interpretability of complex multivariate time series data similarities. The study's key findings are as follows. First, dynamic similarity analysis effectively captured the complexity of building energy use patterns. Second, the approach demonstrated the potential to optimize transfer learning by automatically identifying the most suitable source data. Third, the study explored the feasibility of employing dynamic model selection and ensemble techniques based on temporal similarity changes. This study proposes a practical and scalable methodology to mitigate data scarcity and reduce model development costs, thereby facilitating more efficient, adaptive, and accurate energy demand forecasting.