SCiTransformer: A Trend-Decomposition-Guided Spatio-Temporal Convolution Enhanced iTransformer for Wind Power Forecasting

Wind power, as a vital component of renewable energy, plays a key role in achieving carbon neutrality targets and building a green energy system. However, the highly volatile and non-stationary nature of wind speed presents significant challenges for wind power forecasting. Traditional methods, such as Autoregressive Integrated Moving Average (ARIMA) models and Support Vector Regression (SVR), perform well when modeling linear or weakly nonlinear relationships but often struggle to capture long-term dependencies and non-stationary patterns when faced with multiple overlapping cycles and complex disturbance structures, leading to reduced prediction accuracy and limited adaptability. To address these challenges, this study proposes a wind power forecasting framework that integrates a decomposition mechanism with deep learning strategies, namely the SCiTransformer model. Specifically, the SCiTransformer model applies Seasonal and Trend decomposition using Loess (STL) to explicitly extract trend, seasonal, and residual components from wind power time series data. It further utilizes a Convolutional Neural Network (CNN) module to capture local disturbance features within the residual series, and employs the iTransformer framework to model the global dependencies of trend and seasonal components along the variable dimension, thereby establishing a multi-stage collaborative modeling system tailored to complex sequence characteristics. Empirical studies conducted on four real-world wind farm datasets demonstrate that SCiTransformer consistently outperforms mainstream benchmark modelsincluding Transformer, Informer, DLinear, and PatchTSTin terms of both Mean Absolute Error (MAE) and Mean Squared Error (MSE), achieving superior prediction accuracy and enhanced robustness across different scenarios. Additional ablation experiments further confirm the critical contributions of the STL decomposition and CNN modules to the overall model performance. These findings suggest that the proposed model holds broad potential for applications in multi-scale modeling, non-stationary sequence forecasting, and wind power dispatch prediction.