[Objective] With the high penetration of renewable energy in power systems

coal-fired units are required to provide peak-shaving services with higher frequency and larger depth

making start-up load ramp-up and shutdown load ramp-down more time-critical and cost-sensitive. This study aims to improve the responsiveness and economic performance of non-steady-state operating phases during start-stop transitions under deep peak-shaving conditions

while ensuring compliance with operational safety constraints. [Methods] Multi-month DCS data from a 660 MW coal-fired unit sampled at 5 s were analyzed. Non-steady segments in start-up and shutdown were firstly identified by combining sliding-window range thresholds with steady-state criteria

yielding transient datasets aligned with the underlying operational transitions. Candidate variables were then screened using the maximal information coefficient to retain features most strongly associated with output power

supplemented with a process progress descriptor to represent phase evolution and improve discrimination across different load intervals. To accommodate cross-month operating variability

start-stop events were organized into source and target domains. Within each target-domain segment

a time-stratified split was adopted

with the first 10% used for validation and the remaining 90% reserved for testing

thereby preserving temporal causality and avoiding leakage from future samples. Given the distinct dynamics of start-up load ramp-up and shutdown load ramp-down

two surrogate predictors were developed(LightGBM for tabular feature interactions and a Transformer for temporal dependency representation)

and their hyperparameters were tuned via Bayesian optimization. After model selection

sourcedomain data were combined with the target-domain validation subset for final training. The trained surrogate models were embedded in an NSGA-II multi-objective optimization framework. Under safety constraints and ramp-rate limits

the optimization jointly improved load-ramping capability and reduced fuel consumption

producing executable set-point trajectories and load-change strategies suitable for operational guidance. [Results] The phasespecific surrogate models achieved stronger predictive performance than common baselines such as XGBoost and random forests across operating conditions

and maintained robust generalization in cross-month evaluation. When applied to representative non-steady segments

NSGA-Ⅱ produced feasible set-point trajectories that satisfied the imposed constraints throughout the transient. Relative to baseline operation

the optimized strategies reduced the average duration of non-steady start-stop segments by approximately 5.34% and lowered cumulative fuel consumption by approximately 3.53%

indicating concurrent gains in speed and economy. [Conclusion] A datadriven hybrid modeling and multi-objective evolutionary optimization approach is developed for coal-fired unit start-up and shutdown under deep peak-shaving conditions. By coupling phase-specific power surrogates with constraint-aware optimization

the method generates implementable set-point trajectories that jointly accelerate transient response and reduce fuel use within safe operating envelopes

providing practical support for flexible operation and online decision assistance.