Improve performance for LSTM autoencoder for anomaly detection

Question

I am training on LSTM Autoencoder on time series data for the purpose of reconstruction in order to detect anoamlies in the time series data. One time series dataset contain