Increasing wind farm capacity via overplanting enhances energy production but risks accelerating cable aging if transmission capacity is poorly managed. Consequently, resilient Dynamic Cable Rating (DCR) prediction-defined as the ability to maintain stability under data quality degradation and operational shifts-is crucial for reliable operation. However, achieving this is challenged by limited datasets, missing data, and complex spatio-temporal correlations. To address these issues, a resilient DCR prediction and thermal estimation framework is developed. First, a Conditional Generative Adversarial Network (CGAN) is applied to synthetically augment limited datasets, effectively resolving the load data imbalance. Second, a Spatio-Temporal Graph Attention Residual Shrinkage Network (STGARSN) is proposed. This model integrates an extended Long Short-Term Memory (LSTM) network with Temporal Convolutional Networks (TCN) and a graph attention mechanism to capture complex correlations. Crucially, it incorporates a residual shrinkage module to filter noise and outliers, thereby ensuring model resilience. Finally, to optimize economic performance while minimizing cable aging, a comparative analysis of various overplanting strategies is conducted. Experiments on real cable temperature measurements demonstrate the superior resilience of the proposed model, maintaining high accuracy not only across different forecasting horizons but also under conditions of missing data and sensor noise. The proposed framework accurately predicts DCR and supports long-term offshore wind farm operations through improved economic and technical decision-making.

The effective prediction of dynamic cable ratings (DCR) in the HVDC cable is pivotal for enhancing transmission efficiency and maximizing electricity sales in offshore wind farms. Due to complex wind conditions, traditional machine learning methods, such as support vector machines, struggle to provide accurate long-term DCR predictions and express prediction uncertainties. To address these challenges, this article proposes a novel deep learning framework for dynamic cable rating prediction based on encoder–decoder networks, in which the encoder utilizes Bidirectional extended-long Short-Term Memory networks to encode contextual information from the input data. The decoder introduces an additive attention mechanism, which allows the network to focus on relevant features in the input sequence. In addition, to capture the uncertainty for DCR prediction, a Bayesian neural network approximation method based on the Monte Carlo dropout method is introduced. Finally, this paper introduces a thermal risk estimation method by considering both the maximum conductor temperature limit and the temperature gradient limit. Results demonstrate that the proposed method not only improves electric field distribution but also achieves superior economic benefits.