Process Modeling

This work investigates the use of Naive Bayes models in Federated Learning settings and proposes an FLNB procedure for decentralized structured classification tasks. The motivation comes from application domains such as healthcare, finance and industrial analytics, where models trained from data originating from multiple institutions would be useful, but the centralization of raw data is constrained by legal, ethical or organizational limitations. The proposed procedure combines federated preprocessing with Naive Bayes training. Continuous variables are handled by iterative federated binning, while discrete and categorical values are encoded by deterministic hashing, which provides cross-client representational consistency without requiring a shared plaintext category dictionary. For model aggregation, we introduce the DirSign method, which updates global class-conditional distributions based on correction directions rather than by directly aggregating full local count tables. The method was evaluated on four public structured classification datasets under IID and Dirichlet-based non-IID client partitions. The results show that FLNB-DirSign approaches the performance of the centralized reference model on several datasets while using less directly interpretable local distributional information. The experiments highlight the importance of discretization, smoothing, hash-space size, learning rate and client heterogeneity.

Modern energy infrastructures operate under nonlinear, time-varying conditions and therefore require reliable surrogate models for optimization, predictive control and health monitoring. Existing approaches are brittle: first-principles simulators degrade under parameter drift and imperfect multiscale physics, while purely data-driven methods require large datasets and extrapolate poorly. We propose the Utility-balanced Prior-informed Model (UPriMo), a general training objective that integrates prior knowledge - such as physical laws, topology constraints, operating envelopes and expert heuristics - directly into the loss of any differentiable learner. UPriMo maps each prior residual to a bounded utility and minimizes data misfit plus utility shortfall. This yields automatic trade-offs between data and priors without manual weighting and recovers physics-informed learning as special cases. We demonstrate UPriMo on two energy-relevant problems. First, for stationary laminar pipe flow, a neural surrogate trained on five noisy velocity measurements satisfies Navier-Stokes, boundary and monotonicity constraints, and reduces mean-squared error by an order of magnitude compared with data-driven baselines. Second, for electric water-heater dynamics, nonlinear autoregressive models with exogenous inputs (NARX) augmented with steady-state and monotonicity priors remain stable over 1400-step free runs, whereas unconstrained models diverge. Across both problems, UPriMo improves accuracy, noise robustness and physical plausibility while remaining compatible with standard ML frameworks. Because the utilities are problem-independent, the framework supports a single recipe from component-level surrogates to system-level AI-based forecasting and control in future energy systems.

Accurately forecasting natural water systems is a complex task due to their interconnected structure, where both spatial and temporal dependencies play a critical role. In this work, we applied spatio-temporal graph neural networks of varying complexity to forecast the flow of rivers and the total releases of reservoirs in the Upper Colorado River Basin. Since prolonged droughts driven by climate change can reduce water levels in hydrological systems to critical thresholds, it is essential to forecast to mitigate their negative consequences. The models were trained using five years of historical time series data from directly connected sensor points within a river basin. We evaluated six models and compared their forecasting performance using mean squared error, overall, in boxplots. The graph convolutional recurrent network model performed the best compared to the other five models in the case study, which indicates that the graph convolution with the Chebyshev polynomial has the best forecast accuracy in water system forecasting.