Reliable Operation of Converter Valves in Flexible DC Systems Enabled by Ultra-Short-Term PV Power Forecasting with Intraday Real-Time Correction and Multi-Model Feature Fusion

In flexible DC transmission systems integrating large-scale photovoltaic (PV) power, converter valve reliability is directly threatened by PV output volatility, as PV-induced DC transients overlap with incipient fault features and complicate fault identification. Accurate ultra-short-term PV power forecasting can provide prior operational condition information for converter valve monitoring. However, the forecast irradiance from Numerical Weather Prediction (NWP), which existing forecasting methods rely on, suffers from inherent errors and lacks real-time capability due to its update frequency, failing to meet the accuracy requirements of ultra-short-term forecasting. Furthermore, a single power forecasting model has limited feature capture capability, making it difficult to independently characterize the complex spatiotemporal correlations within PV power sequences. To address this, this study proposes a PV power forecasting method based on intraday real-time information correction and multi-model feature fusion. First, a real-time irradiance correction model based on intraday measured information is designed. It utilizes the Pearson correlation coefficient to match historically similar days and dynamically fuses the measured sequences from these similar days with the current day’s forecast sequence through attention weights, achieving real-time rolling updates of the forecast irradiance. Second, a power forecasting framework featuring “multi-model parallel processing—attention alignment—graph convolutional fusion” is constructed. This framework takes the historical power sequence and the corrected forecast irradiance sequence as inputs, employs four sub-models to extract features in parallel, and performs cross-model feature fusion using a Dynamic Graph Convolutional Network (DGCN), thereby outputting the power forecast for future periods. Experiments show that through the effective utilization of intraday real-time information and the deep fusion of heterogeneous models, this method significantly improves the accuracy and robustness of ultra-short-term PV power forecasting. More accurate future PV power information can provide a dynamic prior operating baseline for converter valves in flexible DC systems, enabling the monitoring system to assess whether observed DC-side transient variations are consistent with expected PV power fluctuations, thereby helping distinguish normal operating transients induced by external PV fluctuations from incipient internal converter valve faults and reducing the uncertainty in operating-state discrimination arising from their overlapping features. Therefore, the proposed method can provide more reliable operating-condition information for converter valve anomaly identification, fault early warning, and operating-state assessment, thereby supporting the reliable operation of converter valves in flexible DC systems with high-penetration PV integration.