Joint Wind and Photovoltaic Power Forecasting with Uncertainty Scenario Generation Based on MS-TCN-GiT
TL;DR
A Multi-Scale Temporal Convolutional Gated iTransformer (MS-TCN-GiT) for joint wind and photovoltaic power forecasting provides accurate point forecasts and compact, interpretable uncertainty scenarios intended for subsequent dispatch.
Abstract
Accurate joint wind and photovoltaic power forecasting is essential for secure operation and dispatch in power systems. Wind and photovoltaic (PV) outputs depend strongly on meteorological conditions and exhibit stochastic fluctuations, multi-scale dynamics, and multivariable coupling. Existing models selectively model the relationships among heterogeneous meteorological variables, but struggle to capture temporal patterns across scales. This paper proposes a Multi-Scale Temporal Convolutional Gated iTransformer (MS-TCN-GiT) for joint wind and photovoltaic power forecasting. Parallel temporal convolutional branches and adaptive weighted fusion (MS-TCN) jointly learn short-term fluctuations and longer-term trends. A gated feature-selection mechanism (GiT) is embedded in the iTransformer variable-attention framework to screen and reweight meteorological variables dynamically. Point forecasts are combined with an error-statistics-based center-trajectory method that generates low-, moderate-, and high-output scenarios. Experiments on two years of State Grid microgrid data show that MS-TCN-GiT outperformed all evaluated baselines. Relative to TCN, it reduced capacity-weighted MAE and RMSE by approximately 23.2% and 22.2%, respectively, while increasing R2 to 0.9373. The framework therefore provides accurate point forecasts and compact, interpretable uncertainty scenarios intended for subsequent dispatch.