As photovoltaics (PV) scale rapidly to meet climate and energy goals, ensuring long-term reliability and performance is increasingly important. With global installed PV capacity surpassing the terawatt scale, and expected to continue to grow exponentially in the coming years, even modest improvements in system lifetime, monitoring, and degradation understanding can yield major environmental and economic gains. At this scale, where millions of systems must operate reliably for decades, standardised diagnostics and fleet-wide assessment frameworks are essential. This thesis addresses the challenge of quantifying PV degradation in real-world conditions by developing and validating a set of data-driven methods that improve both long-term performance loss assessments and short-term fault detection. Though demonstrated mainly on Swiss systems, the approaches are broadly applicable across climates, technologies, and deployment contexts, offering tools for more accurate benchmarking, maintenance, and optimisation as PV continues its expansion in the terawatt age.

The first part focuses on improving long-term performance assessments while reducing uncertainty. A key contribution is the novel multi-annual Year-on-Year (multi-YoY) method, which increases comparison depth and reduces statistical noise in performance loss rate (PLR) estimates. Applied to synthetic and real-world datasets, it lowers statistical uncertainty by over 90% compared to conventional methods, enabling clearer benchmarking across systems of varying age, quality, and environmental exposure.

The second part addresses short-term variability and its impact on long-term diagnostics. It introduces a fault detection and diagnosis algorithm (FDDA) that identifies reversible loss conditions such as snow, shading, and inverter downtime, based on deviations from expected output. Applied to over 300 PV strings, the FDDA classifies six fault types at 15-minute resolution and enables the definition of intrinsic PLR (i-PLR), which excludes temporary faults and isolates true degradation. In one case, fault filtering reduced the apparent degradation by 80%, linked to a threefold increase in partial shading within a few years due to a growing tree. Across the fleet of PV strings, i-PLR was significantly lower in fault-prone systems, and differences correlated strongly with the fault time factor, a new metric quantifying fault occurrence. Overall, the FDDA reveals how transient faults can distort PLR trends.

The final part translates the research into an integrated monitoring platform for PV operators, developed and deployed in collaboration with 3S Swiss Solar Solutions. The platform combines fault detection and long-term analytics into a scalable, modular tool that supports automated diagnostics and predictive maintenance. Real-world case studies include inverter failure detection, soiling loss tracking and recovery leading to 25% yield gains, and shading impact quantification.

Overall, this thesis bridges the gap between short-term fault detection and long-term degradation analysis. The tools and methods proposed enable more accurate health assessments, improve maintenance and warranty planning, and support the design of more reliable PV systems. These contributions help advance reliable, data-driven PV operation at terawatt scale, supporting global energy and climate goals.