A new self-supervised model learns to forecast energy demand and solar output without needing labeled data for every building or panel it studies.
Researchers built a distributed Joint Embedding Predictive Architecture (JEPA), a self-supervised approach that predicts hidden representations of masked chunks of energy time-series data instead of raw values. The team trained it on energy consumption and generation datasets, including scenarios with missing data, and benchmarked it against a Transformer forecasting model. On building energy consumption data, JEPA matched the Transformer's accuracy and beat it in three of five consumer clusters. On solar generation data from panels the model had never seen before, it outperformed the baseline on nine of ten sites, with R^2 scores of 0.73 to 0.88 versus under 0.45 for the Transformer, and held up better when data went missing.
Energy forecasting today is usually a patchwork of site-specific models, one per building or solar farm, because assets differ too much for a single model to generalize well. This work suggests a single self-supervised model can learn transferable patterns across a fleet of buildings and panels. That matters for utilities and grid operators trying to forecast output from thousands of new distributed solar installations without training a bespoke model for each one.
This is a single arXiv preprint tested on a limited set of assets, not a deployed grid tool, so the real test is whether these gains hold up on messier, larger-scale data.