Abstract
Abstract
Renewable-aware manufacturing requires forecasts that remain informative across operational horizons while respecting the different physical semantics of grid demand, renewable generation, building consumption, and industrial loads. We present NEO-Forecast, an evidence-controlled cross-domain forecasting methodology that separates entity generalization, domain transfer, and semantic/task transfer instead of treating heterogeneous normalized series as interchangeable. The framework combines leakage-safe chronological evaluation, domain/entity context, horizon-specific prediction, and an explicit compatibility layer intended to test when shared temporal knowledge is beneficial and when it produces negative transfer. Its engineering boundary is a forecast-to-decision interface that maps predictions to constrained manufacturing-energy actions without conflating forecasting with production optimization. The current research record reports aggregate MAE/RMSE/MAPE/error-variance values of 5.9/9.8/6.8%/2.6 for NEO-Forecast versus 8.2/13.4/9.6%/3.7 for LSTM, together with a separate normalized-RMSE record of 0.089 versus 0.112 for a Transformer and 0.145 for XGBoost. These aggregate values are retained only at their supported evidence level; dataset-, horizon-, transfer-, uncertainty-, and manufacturing-specific superiority is not inferred without corresponding artifacts. The revised study therefore defines a falsifiable experimental programme for compatibility-aware transfer, horizon-dependent transfer benefit, negative-transfer controls, and downstream manufacturing utility, establishing a reproducible route from forecasting evidence to engineering claims.