Hybrid Prophet–TCN Framework for Deterministic and Probabilistic Wind-Power Forecasting

Accurate wind-power forecasting is essential for maintaining grid stability under high renewable penetration. This paper presents a hybrid Prophet–Temporal Convolutional Network (TCN) model that combines structural time-series decomposition with deep nonlinear temporal modeling for turbine-level point and probabilistic forecasting. Prophet captures long-term trend and weekly, monthly, yearly seasonalities with meteorological regressors, while the TCN learns short-term multivariate dependencies from 30-day windows using dilated causal convolutions with residual and normalized blocks. A lightweight attention layer adaptively weights the two branch outputs to produce the final hybrid forecast. The model is evaluated on a four-year dataset from Jeju Island, South Korea, comprising 1,461 daily observations from five turbines and 42 meteorological variables. The hybrid forecaster achieves seasonal R2 values up to 0.9483 and monthly MAE and RMSE as low as 25.8 MWh and 37.8 MWh. Daily and seasonal analyses show R2 generally above 0.86, with performance degrading only under the most volatile winter conditions. Probabilistic evaluation using CRPS, pinball loss at τ =0.1,0.5,0.9, and prediction-interval coverage at 80% and 90% indicates well-calibrated uncertainty, with CRPS closely aligned with MAE and coverage typically within 3–4 percentage points of nominal levels. Overall, the proposed Prophet–TCN framework provides accurate, interpretable, and operationally robust wind-power forecasts suitable for risk-aware grid operation and energy planning.