This paper presents a latent-space dynamic neural network (LSDNN) model for the multi-step-ahead prediction and fault detection of a geothermal power plant’s operation. The model was trained to learn the dynamics of the power generation process from multivariate time-series data and the effects of exogenous variables, such as control adjustment and ambient temperature. In the LSDNN model, an encoder–decoder architecture was designed to capture cross-correlation among different measured variables. In addition, a latent space dynamic structure was proposed to propagate the dynamics in the latent space to enable prediction. The prediction power of the LSDNN was utilized for monitoring a geothermal power plant and detecting abnormal events. The model was integrated with principal component analysis (PCA)-based process monitoring techniques to develop a fault-detection procedure. The performance of the proposed LSDNN model and fault detection approach was demonstrated using field data collected from a geothermal power plant.
Accurate characterization and prediction of subsurface flow systems rely on assimilating dynamic monitoring and performance data into high-dimensional nonlinear physical models. Conventional ensemble-based data assimilation methods rely on finite-sample covariance estimates and may require large ensembles or repeated assimilation to reduce sampling errors and address nonlinear observation relations. Deep-learning-based latent representations can reduce dimensionality and computational cost, but dimensionality reduction alone does not necessarily simplify the nonlinear parameter--observation relationships that govern Bayesian model updating. We develop a structured deep latent state-space model specifically designed for data assimilation. The proposed linearized latent-space formulation uses explicit linear latent transition and observation operators, while variational state and parameter encoders regularize their latent representations toward Gaussian distributions. Two assimilation-informed objectives, a filtering loss and a posterior-trajectory loss, promote latent representations better suited for inversion and enable direct Kalman filtering. The framework is evaluated using synthetic geological carbon storage experiments and compared systematically with ensemble-based data assimilation. Results show that assimilation-informed learning improves posterior model calibration and prediction, while direct Kalman filtering eliminates finite-ensemble covariance estimation and repeated assimilation, substantially reducing online computational cost. The proposed framework provides an efficient approach for sequential calibration of high-dimensional subsurface flow models.