Table 8 Comparison with AI methods.
From: Bayesian-optimized LSTM-DWT approach for reliable fault detection in MMC-based HVDC systems
Protection method | Detection-method | Feature selection | Noise immunity (dB) | Fault resistance withstanding ability | Sampling-rate | Pre-training settings | Complexity in training setup time | Other remarks | Average-recognition rate |
|---|---|---|---|---|---|---|---|---|---|
Ref.37 | ANN | DWT | High (18.0 dB) | High (350.0 Ω) | Medium | Trial and error | Medium | A generalized linear model to set up hyperparameters of a non-linear model causes miss-convergence at a small window length | 90.61% |
Ref.16 | ANN | Voltages | Low | Low (50.0 Ω) | Low | Trial and error | High | Due to the large window length, response time is slow | 94.35% |
Ref.38 | SVM | WT | Not mentioned | Low | High | Not mentioned | High | Do not discuss noise immunity in the complicated and dynamic environment of the MT-HVDC networks | 92.92% |
Ref.39 | ANN | Fast FT | Low (50.0 dB) | Low (100.0 Ω) | Medium | Trial and error | Medium | An in-depth analysis of noise immunity is missing, and the fault resistance range is -0.01–100-Ω- | 88.85% |
Proposed methodology | Improved LSTM | Norm | High (20.0 dB) | High (480.0 Ω) | Medium | Optimal settings (BO) | Low | Due to the optimal training solution, it is less prone to miss-convergence at a small window length | 99.04% |
