Adaptive-Resolution Long-Sequence Framework for Counterfactual Building Retrofit Energy-Savings Forecasting

Abstract

This paper proposes an adaptive-resolution long-sequence framework for counterfactual building retrofit assessment that preserves short operational transitions while retaining months of weather, occupancy, and control context Multivariate meter, weather, occupancy, equipment, and intervention descriptors are encoded as ordered engineering tokens. A learned resolution controller preserves fine tokens near thermal and operational transitions, protects intervention and calendar boundaries, and merges redundant intervals elsewhere. A retrofit-conditioned, duration-aware Hyena operator then mixes the compressed long sequence, while matched decoders estimate post-retrofit demand and the corresponding no-retrofit trajectory. On an illustrative and replaceable benchmark containing 42 commercial buildings and 5.21 million 15-min observations, the model obtains 5.8% normalized mean absolute error, 7.4% normalized root-mean-square error, and 3.6 percentage-point error in annual saving rate. Relative to a dense Transformer, error falls by 18.3%, peak-direction accuracy rises from 85.6% to 91.2%, inference memory decreases by 38.6%, and one-building annual evaluation requires 0.84 s. The contribution is therefore an integrated engineering formulation of temporal allocation, long-context mixing, and retrofit attribution rather than a direct transfer of the original biological-sequence model. Because the benchmark and reported values are illustrative, all empirical outputs must be replaced or independently verified before submission.