High-Dimensional Feature Adaptive Selection and Season-Specific Short-Term Photovoltaic Power Forecast

Accurate prediction of photovoltaic (PV) power is crucial for ensuring the stable operation of power grids integrated with PV systems. However, the complex coupling between environmental variables collected by various sensors and PV power increases the uncertainty of prediction tasks. To overcome this challenge, this paper proposes a high-dimensional feature adaptive selection and season-specific short-term PV power forecasting method (FASS-DCI). Firstly, the method designs a feature utility evaluation index called Feature Bottleneck Score (FBS) to assess the complex coupled data from the environmental measurement equipment at the PV station, and uses this score to select the optimal feature combination. Secondly, to address the impact of seasonal changes on PV systems, a seasonal positional encoding based on the cosine and sine amplitude-phase characteristics is proposed to enhance the model’s seasonal awareness. Finally, the method combines dilated causal convolution with the Informer architecture and introduces high-dimensional features to achieve the prediction task. We selected data from sensors in an actual PV power plant environment, covering four seasons and two types of weather, resulting in eight different environmental conditions for model validation. The experimental results indicate that, compared to current popular prediction algorithms, the proposed method improves RMSE, MAE, R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> and metrics by 30.17%, 30.41%, and 4.24%, respectively.