A trustworthy disruption prediction framework on HL-3: integrating OOD detection for reliability and real-time SHAP for interpretability

Plasma disruption poses a severe challenge to the safe operation of tokamak devices and future fusion reactors. Although deep learning has demonstrated outstanding accuracy in disruption prediction, its "black-box" decision-making nature and "blind confidence" when facing out-of-distribution (OOD) data raise serious concerns about whether such methods can be safely and reliably deployed in the control systems of future fusion reactors. To address these issues, this study proposes a highly trustworthy disruption prediction framework based on Temporal Convolutional Networks (TCN) for the HL-3 tokamak, constructing a dual safety barrier of "knowing boundaries and understanding mechanisms" by integrating OOD detection with real-time interpretability analysis. First, the framework introduces autoencoder-based manifold learning and data generation techniques to delineate the model's safe cognitive boundary. This mechanism can accurately identify unseen operating conditions or anomalous diagnostics that deviate from the training distribution, effectively intercepting random outputs and false alarms produced by the model in unknown domains. Second, under the premise that input data is legitimate and within the reliable boundary (in-distribution), for high-confidence disruption predictions, a Monte Carlo-accelerated SHAP (SHapley Additive exPlanations) method is employed for real-time feature attribution. This algorithm can quantify the contribution of each diagnostic channel to the prediction result within the time window required by the control system, revealing the key physical causes of instability onset (such as horizontal displacement, impurity accumulation, etc.) in real time. Validation on the HL-3 dataset demonstrates that the OOD detection module achieves an average area under the receiver operating characteristic curve (AUC) of 0.931 across seven anomaly modalities, effectively intercepting OOD inputs and reducing false alarms; the Monte Carlo SHAP method completes per-window feature attribution in 4.72 ms (T=100 samples) and 1.50 ms (T=10 samples), representing a 5.2x and 16.5x speedup over the exact SHAP computation (24.72 ms), while satisfying the real-time requirements of the HL-3 plasma control system. This trustworthy framework provides real-time decision support with clear physical significance for operational control, establishing an important paradigm for the safe and reliable application of deep learning in future large-scale fusion devices.