FCC-Bench and Adaptive Multi-Scale Inference for Multivariate Forecasting Using Data from a Continuous FCC Unit

Fluid catalytic cracking (FCC) units are central to modern oil refining. FCC operation is complex, with strong nonlinearity and high-dimensional spatiotemporal coupling, making process monitoring, soft sensing, and operational optimization difficult. Multivariate forecasting supports these tasks. Public datasets and forecasting benchmarks provide limited data for one continuous FCC process, and forecasters transferred to FCC operation often need retraining to maintain accuracy as operation changes. To address these issues, we construct the Fluid Catalytic Cracking Benchmark (FCC-Bench) and propose Adaptive Multi-Scale Inference (AMSI). FCC-Bench focuses on one continuous FCC process, with 129 variables over 62.5 days, capturing operating changes and supporting traceable saved-array metric calculation. AMSI runs a frozen forecaster at three context lengths and fuses forecasts with adaptive per-variable weights, so short contexts respond quickly to sudden changes while long contexts reduce noise during slow changes. In the FCC-Bench last-20% rolling comparison, AMSI reaches the lowest mean MSE and MAE among the evaluated methods of 1.2909 and 0.8101. The same settings still give the lowest error on another period from the same unit and on another unit.