Short-term photovoltaic power forecasting under complex meteorological conditions: a physics-driven classification and hybrid deep learning framework

Existing classification methods often suffer from insufficient physical interpretability, and single prediction models exhibit weak generalization capabilities under complex scenarios. This paper proposes a short-term photovoltaic power forecasting (STPVPF) method based on a physically driven classification and hybrid deep learning framework. First, a Fluctuation Energy Loss Index (FELI) is proposed to quantify the degree of volatility, and four fluctuation scenario labels with explicit physical significance are established by combining the Jenks natural breaks algorithm. Second, the Hilbert-Huang Transform (HHT) is introduced to extract the nonlinear time–frequency features of Numerical Weather Prediction (NWP) data. A ShuffleNet module embedded with a Parallel Multi-dimensional Attention (PMDA) mechanism is then utilized to achieve high-precision identification of the target day’s scenario. Subsequently, a novel hybrid architecture, ResNet-CTformer, is constructed. This architecture effectively integrates the residual learning of ResNet, the local feature extraction of CNNs, and the global temporal modeling advantages of Transformers, thereby overcoming the bottleneck of network degradation and enhancing the robustness of long-sequence forecasting. Finally, an empirical study based on a PV power station in Jilin Province demonstrates that the proposed method possesses significant advantages in accuracy, with NRMSE and NMAE reduced by 16.91% and 14.59% on average compared to existing mainstream methods.