Liana Toderean and colleagues report that Distributed JEPA, an AI system that learns patterns in energy measurements, beat a forecasting baseline at nine of ten held-out solar sites. The September 15 preprint tests whether learned patterns transfer between installations.
The system starts with TS2Vec, an existing method for turning a time series, a sequence of measurements, into numerical features. Its original researchers train it by comparing altered views of the same sequence, encouraging matching observations to receive consistent representations. Each timestamp can therefore carry information about its surrounding context. Those features can support forecasting or anomaly detection through a separate prediction model. This separation gives developers two components to examine: whether the features preserve useful patterns, and whether the final prediction model can turn those patterns into the quantity needed.
Distributed JEPA adds a Joint Embedding Predictive Architecture: it combines measurement features and site information, predicts features for hidden periods, then decodes them into energy values. Training penalties discourage uninformative, nearly identical predictions. Local encoders allow raw measurements to stay at their originating sites.
The solar comparison used 40 training sites and ten held-out sites, with 168 hours of context predicting the next 24 hours. The baseline was a Transformer, a neural network that relates observations across a sequence. Both approaches used TS2Vec features. Crucially, both adapted their output components using 70% of each held-out site's data before evaluation. This was transfer with local adaptation, rather than an immediate forecast at a new installation.
One exception exposes the limits of that transfer. At a low-variation site, JEPA's R-squared score was -2.815 against 0.276 for the baseline. R-squared compares squared prediction errors with errors from predicting the measured series' mean. A perfect fit scores one; zero corresponds to that constant reference; negative values indicate larger errors. It is not a percentage of forecasts that are correct. When measurements vary little, the reference error is small, so this score should be read alongside absolute errors. JEPA also lost that site's absolute-error comparison.
The broader building test was mixed: the baseline had lower mean absolute error in all five groups. For an operator considering this approach, the useful next test would preserve the local adaptation step and report its data and computing cost. It should also separate nearly flat production profiles from strongly varying ones, so a good overall result cannot conceal installations where forecasts deteriorate. Independent reproduction and comparison with additional forecasting methods would help establish when the extra learning stage earns its place in an operational system.
Sources
- Liana Toderean and colleagues, Distributed JEPA, September 15, 2026, arXiv:2609.17029v1. Methods and Table 4; preprint. Research paraphrased and analyzed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).
- Zhihan Yue and colleagues, TS2Vec: Towards Universal Representation of Time Series, AAAI 2022. Background on timestamp representations and downstream prediction.
- Illustrative photograph, not a study installation: Solar Panels on Rooftop, by Esthee2010, CC BY-SA 4.0. Original photograph unaltered; illustrative setting.
