Strategies for maintaining inverter temperature within operational limits to enhance longevity: A comprehensive literature review

Abstract

• PRISMA-based systematic review of machine learning for PV inverter reliability. • Critical analysis of ML methods for fault diagnosis, junction temperature estimation, and ancillary control. • High reported accuracy in diagnostics, but limited validation under real operating conditions. • Data-driven temperature estimation improves monitoring but shows limited generation across systems. • Key gaps identified in standardisation, scalability, and real-time implementation of ML approaches. The reliability of in photovoltaic (PV) inverters remains a critical limitation in modern power systems, with thermal stress at semiconductor junction level identified as a primary driver of degradation and failure. While machine learning (ML) has increasingly been proposed to enhance inverter reliability, its actual contribution, scope, and limitations remain insufficiently synthesised in the literature. This paper presents a systematic review of ML-based approaches for PV inverter reliability, focusing on fault diagnosis, junction temperature estimation, and ancillary service control. Following the PRISMA framework, relevant studies are systematically identified and critically analysed. The review shows that ML techniques have achieved high performance in fault detection under controlled conditions, yet their robustness under varying operating environments and real-world noise remains inadequately validated. In junction temperature estimation, data-driven and hybrid models demonstrate improved accuracy, however, their dependence on high-quality datasets and limited transferability across inverter topologies constraint practical deployment. In the context of ancillary service, ML-driven control strategies exhibit potential to influence inverter loading and thermal profiles, but their integration with reliability-oriented control remains fragmented. Overall, the literature reveals a strong methodological focus on model development, with limited attention to standardization, scalability, and real-time implementation. The paper identifies critical research gaps and emphasises the need for unified, thermal-aware, and deployment-ready frameworks to advance ML-based PV inverter reliability in practical applications.