The widespread deployment of Internet of Things (IoT) networks has actualized omnipresent device interconnectivity. Despite technological advancements, IoT edge devices suffer persistent energy bottlenecks from suboptimal coordination of power acquisition and adaptive management. Self-charging power sources (SCPS) aim to achieve autonomous operation through monolithic integration of three core components: energy harvesters, power management circuits, and supercapacitors/batteries. These devices enable continuous ambient energy harvesting, providing uninterrupted power supply for wearable electronics and IoT applications. Nevertheless, material selection and component design remain key challenges in SCPS development. As an essential artificial intelligence paradigm, machine learning (ML) enables data-driven material and structural design based on historical experimental datasets, thereby elevating SCPS performance to superior level. This paper reviews the development of SCPSs and the application of ML in SCPSs, with a particular focus on SCPSs with triboelectric nanogenerators (TENGs) and supercapacitors (SCs). A generalized ML workflow with suggested parameters is proposed to guide the performance prediction of TENG by incorporating previous theoretical research. Additionally, ML-guided design of carbon-based SC materials and computer-aided suppression of self-discharge performance are selected as typical examples to discuss. The combination of ML and SCPS is expected to push forward more efficient and self-sufficient IoT applications.