A method for anomaly detection in cascade hydropower station water level data based on variational autoencoders

The current methods for anomaly detection in water level data are often limited in accuracy and application scope, failing to fully consider the non-linearity and non-stationarity of cascade hydropower station water level data. This paper proposes an improved water level data anomaly detection method based on Variational Autoencoders (VAE). This method integrates Bidirectional Long Short-Term Memory networks (Bi-LSTM) with the foundational VAE framework to learn the forward and backward temporal correlations in water level time series, enabling high-precision data reconstruction. To adapt to the varying water levels in different periods at cascade hydropower stations and enhance the sensitivity of anomaly detection, dynamic setting of anomaly thresholds is achieved by combining Exponential Weighted Moving Average (EWMA) with a sliding window technique. Compared to methods like ARIMA, Isolation Forest, and VAE-LSTM, this proposed approach demonstrates more stable detection performance with higher overall recall and F1 scores when dealing with various types of water level anomalies.