Photovoltaic power forecasting and anomaly detection: a comparative study of statistical, deep learning, and zero-shot foundation models on seven years of real plant data

Abstract

Faults and gradual performance degradation in photovoltaic plants often remain undetected, causing energy losses and increasing maintenance costs. Selecting an appropriate forecasting model for residual-based anomaly detection requires understanding the relative contributions of model architecture and sensor availability; a question not yet addressed by systematic benchmarks on long-term real data. This paper presents a controlled two-phase benchmark based on seven years of real 15-minute PV production data from a grid-connected plant with two synchronised inverters. Four models are evaluated: a multiplicative Prophet model, a Seasonal Autoregressive Integrated Moving Average model with exogenous regressors, a bidirectional Long Short-Term Memory network, and Chronos-2, a zero-shot foundation model requiring no fine-tuning on plant-specific data. In Phase 1, models use only historical production data. In Phase 2, plane-of-array irradiance and module temperature are added as real-time covariates. With covariates, all models exceed R2 = 0.97, enabling a shared three-stage anomaly detection pipeline to identify five operationally distinct fault categories, including inverter blackouts, partial faults, and snow accumulation. Across the test year, two confirmed fault events are identified: a total blackout with an estimated energy loss of 80 kWh and a partial fault with a loss of 21 kWh; the blackout is detected by all models. Zero-shot spatial transfer of the Prophet model to the second inverter achieves R2 = 0.966, outperforming full retraining. Results show that covariate availability has a greater impact on forecast accuracy than model architecture itself. In covariate-rich settings, higher forecast accuracy corresponds to reduced anomaly detection sensitivity. The findings demonstrate that zero-shot foundation models are ready for deployment in PV plants lacking historical data, enabling scalable intelligent monitoring without upfront training.