Time Series Prediction for Anomalies Detection in Concentrating Solar Power Plants Using Long Short-Term Memory Networks

Concentrating Solar Power (CSP) plants that use a parabolic trough system rely on Heat Transfer Fluid (HTF) to absorb thermal energy from sunlight. The heated HTF is then used in thermal power blocks to produce electricity in conventional steam generators. Unexpectedly high HTF temperatures may lead to degradation of the system components and reduced efficiency. Therefore, closely monitoring and maintaining the HTF’s operational temperatures is crucial to ensure the system’s efficiency and longevity. This paper focuses on the detection of over-temperature anomalies of HTF in CSP plants. Encoder-decoder Long Short-Term Memory (LSTM) networks are applied to predict HTF temperature in a time series, and subsequently, anomalies are detected based on the mean average error threshold. The study concludes by analysing the effectiveness of the encoder-decoder LSTM-based method in detecting over-temperature anomalies in historical plant data. The proposed approach allows operators to take preventive measures before any potential alarms by providing a 300-s forecast window.