Predicting the Temperature Increase Trend of a Generator Using RNN, GRU, and LSTM Algorithms at Nam Ngum 1 Hydropower Plant in Laos

Abstract

Synchronous generators are integral to the operation of hydropower plants. Stator faults, including short circuits, open circuits, and inter-turn faults, can cause severe performance issues and even catastrophic failures if not identified and mitigated promptly. Traditional generator monitoring methods such as periodic inspections and time-based maintenance often fail to detect subtle or evolving faults. This study proposes an advanced predictive maintenance approach utilizing deep learning techniques to monitor generator health at Nam Ngum-1 (NNG-1) Hydropower Plant in the Lao People’s Democratic Republic (PDR). Accurate temperature forecasting is vital for predictive maintenance, as excessive heat can lead to performance degradation and costly downtime. Using time-series data from the plant’s supervisory control and data acquisition (SCADA) system, including parameters such as power output, voltage, current, and cooling system temperatures, this research evaluates the models’ ability to capture temporal dependencies critical for precise trend prediction. Among the models tested, the results demonstrate that LSTM, with one hidden layer, achieved the highest accuracy based on MSE, RMSE, MAE, and R-squared, outperforming GRU, which had an R-squared of 98.60%, and RNN, which achieved 97.04%. When implemented with two hidden layers, LSTM maintained its superior performance with an R-squared of 98.34%, compared to GRU's 97.93% and RNN's 92.68%. These results demonstrate LSTM's exceptional capability in capturing both short- and long-term temperature dependencies, making it particularly suitable for predictive maintenance applications. The model's high accuracy in temperature forecasting enables early fault detection, helping to prevent performance degradation and reduce costly downtime in hydropower operations.