A GCN-LSTM multi-task learning framework for fault prediction and reliability management of hydropower generating units

Abstract

Fault prediction and reliability management are essential for ensuring the safe and stable operation of hydropower generating units under variable hydraulic and electrical operating conditions. Traditional threshold-based monitoring and periodic maintenance strategies are insufficient to capture the nonlinear coupling relationships among key components and the long-term degradation characteristics of hydropower units. To address this problem, this paper proposes a graph convolutional network and long short-term memory (GCN-LSTM) based multi-task learning framework for fault prediction and reliability management of hydropower generating units. Fault tree analysis (FTA) and the analytic hierarchy process (AHP) are first used to identify critical fault modes and construct a multi-source monitoring variable system covering hydraulic, mechanical, electrical, thermal, and lubrication parameters. A hybrid adjacency matrix is then established by combining physical mechanism connections with data-driven correlation relationships. Based on this matrix, the GCN extracts spatial coupling features and suppresses redundant noise, while the LSTM models temporal degradation under changing water head, guide vane opening, and unit load. A multi-task learning structure simultaneously predicts the remaining useful life (RUL) and diagnoses fault categories. The predicted RUL, fault probability and component importance are integrated into a dynamic reliability index and a maintenance priority strategy. Experiments on a 600 MW hydropower unit dataset show that the proposed framework reduces the RUL prediction RMSE by 28.3% relative to a standalone LSTM and improves fault diagnosis accuracy to 96.4%, providing quantitative support for reliability-centered maintenance.