Accurate forecasting of emergency department (ED) patient flow is essential for safe staffing, capacity planning, and escalation protocols; however, demand demonstrates significant diurnal and weekly fluctuations, as well as disruptions due to holidays and weather, and feedback from inpatient boarding. We compare multi-horizon forecasts (1h, 4h, 12h, 24h) from seasonal ARIMA and Prophet, tree-based learners (Random Forest, XGBoost), sequence models (LSTM), and a Temporal Fusion Transformer (TFT). We assess point accuracy (MAPE, sMAPE, MAE, RMSE), calibration, residual structure, and operational relevance using a six-year synthetic yet realistic dataset that includes exogenous signals such as calendar, weather, syndromic indicators, occupancy, and triage mix. Results indicate that horizon-specific training with exogenous features produces significant improvements; TFT achieves the highest overall accuracy and calibration, aligning with recent multi-site and occupancy-forecasting studies that highlight exogenous factors and probabilistic assessment (Rostami-Tabar et al., 2023; Porto et al., 2024; Tuominen et al., 2024). Long-term accuracy goes down, but XGBoost is still competitive at 12–24 hours. Using exceedance probabilities, forecast envelopes, and interpretable attributions (permutation importance, attention weights), we turn forecasts into operations that take risks into account. Figures 1–4 and Tables 1–2 present decision-relevant facts instead of comprehensive statistics, in accordance with the directive to convey a coherent operational narrative. The results support a layered deployment: short-term tactical control with TFT/LSTM, mid-term scheduling with XGBoost, and strong classical fallbacks. We finish with a deployment playbook that includes data quality, drift monitoring, bias audits, and change management to make sure that everyone can use it safely and fairly. Recent public health signals, such as increased influenza seasons, further necessitate strong exogenous coverage in production pipelines (CDC, 2024).