Universal metro flow prediction models suffer from performance degradation or even worse accuracy than naive baselines at low-signal stations, because stations differ in function of stations and the distinct characteristics of passenger flow patterns. To solve this model selection problem, we develop an adaptive forecasting framework to define the applicability boundaries of intra-day mode decomposition (L3: mode-specific LightGBM with mode-feature interactions) in heterogeneous station environments. Based on 24-day data of 34 stations on Metro Line B, we use K-means to cluster five-dimensional operational features, dividing stations into three types: 2 hub/commercial stations with bimodal flow, 7 commuter-dominant stations with tidal flow, and 25 suburban/residential stations with poor flow regularity. L3 distinguishes weekday commute and weekend leisure patterns via clustering of daily flow curves and constructs mode-feature interaction terms in a three-tier feature framework. We compare L3 with L2 (standard LightGBM) and L1 (historical mean at the same HH: MM) across 60/120/180-min horizons, finding heterogeneous model performance: L3 reduces hub station errors by 37.4% at 60 min and retains a 22.5% advantage at 180 min; it brings 37.7% short-term improvement to Type 3 stations with <1% incremental gain over L2; it causes a 9.6% performance drop for suburban stations in long-term forecasts, showing overfitting risks of forced decomposition. Overall, this paper constructs a multi-layer decision-making mechanism based on station type, passenger flow scale, and prediction time horizon, and proposes quantitative thresholds (weekend elasticity of 1.0, 500 passengers per 10-min slot, and forecast horizon of 120 min), which provide more accurate and adaptive models for metro operation.