Fault diagnosis in photovoltaic power plants via ARX residual augmentation

Online fault detection and diagnosis (FDD) in photovoltaic (PV) plants must identify rare faults in imbalanced monitoring streams without using future information. We present a model-informed pipeline in which a multivariate autoregressive model with exogenous inputs (ARX) is calibrated on normal-only data and then frozen as a data-driven healthy-regime reference. Its one-step voltage predictions and signed residuals augment irradiance, temperature, and current measurements for five-class classification. The proposed ARX–NN processes this representation causally at each sample through a compact shared encoder with fault-classification and auxiliary voltage-estimation outputs. All classifiers are evaluated on identical chronological boundaries comprising an 80%/20% holdout and four expanding future-contiguous blocks. On the chronological holdout, ARX–NN attains macro F 1 = 0.788 ± 0.046 , compared with 0.632 ± 0.010 for NN using operating inputs and 0.650 ± 0.039 when the ARX channels are masked; this ordering persists across all four later blocks. XGBoost using the same ARX-augmented representation is the benchmark accuracy leader at 0.912 , compared with 0.542 using operating inputs. These results show that the complete ARX-augmented feature bundle is informative across classifier families, while positioning ARX–NN as a causal joint-output neural implementation rather than a universal replacement for XGBoost. At the selected event operating point, ARX–NN has shorter total false-alarm duration but more false-alarm episodes than XGBoost, indicating a distinct alarm profile. Because the benchmark lacks explicit timestamps and an external-plant test, conclusions are limited to later contiguous rows from the ordered stream.