Physics Based Models in Simulation and modeling

Digital Twin analytic models are derived from and validated against vast volumes of time series data from GE’s install
base. The models, used by GE business applications, are designed to predict an array of operating conditions in a power
plant environment. While GE employs a wide variety of analytic models for operations efficiency, the following describe those
most impactful that contribute to the value of the GE Digital Twin.
• Plant and asset thermodynamic model (Thermal)
• Anomaly models and detection methods (Anomaly)
• Life models (Lifing)
• Dynamic Estimation and Model Tuning (Transient)
• Flow and Combustion Models (Transient)
4.1.1 Plant Thermodynamics Models
Plant Thermodynamic Models predict plant performance under different operating conditions, dispatch modes and grid (or
customer) requirements both under steady state as well as transient operation. The model uses GE gas turbine power plant
design knowledge, including GT, ST, and HRSG, with advanced computational methods to accelerate the execution time and
enables real time decision-making. With GE’s deep expertise in combined cycle gas turbine design and the wealth of data
gathered GE large fleet of fielded gas turbine plants, the model robustness, accuracy, and agility is unprecedented. GE is
currently using GateCycle as our source for a core Digital Twin thermal application. Specifically, GE is using the heat-balance
engine within GateCycle (“HBE”) and developing an improved version, HBE7.
The model is created using a user interface (UI) that allows an engineer to define each plant component and how those
components are connected to each other. The heat-balance engine contains the fundamental physics of each component
in the power plant as well as a “solver” that controls how these components interact such that the final result complies with
the laws and principles of thermodynamics. This creates the as-designed model of the power plant.
Once a model is created, the model can be used in multiple ways. During commissioning, the as-designed model is critical
as this enables the test engineers to confirm whether the plant meets acceptance criteria for thermal efficiency and
output. If there are shortcomings, the model helps the engineer identify the root causes and accelerates remediation. After
commissioning is complete, the tuned model represents the as-running condition of the power plant. Connecting this tuned
model to a Monitoring and Diagnostics infrastructure, allows the plant to be monitored both by the customer and by GE.
The tuned thermal model can also be used within the control system. Building on the work done in creating closed-loop
optimal control (CLOC), GE has been able to demonstrate improved regulation performance, optimized load generation
balance, and elimination of over generation costs. This can be taken to the next level with advances in signal processing and
model-based control combined with GE’s DCS-OEM-agnostic APAL platform.
4.1.2 Anomaly Models and Detection Methods
Every day, GE Power collects more than 44,000 operating hours of data from thousands of globally deployed gas turbines,
steam turbines and generator assets. Remote condition monitoring and anomaly detection of a power generation asset
involves the full spectrum of data collection, data processing, mechanical condition monitoring algorithms, and alarm

GE Digital Twin: Analytic Engine for the Digital Power Plant © 2016 General Electric Company. All rights reserved. 16
disposition, diagnostics and recommendations for improvement. The purpose is to monitor the operation status of each
asset by using time series data collected and transmitted from the customer site to detect abnormality at near real-time.
Generally, there are two types of abnormalities: (1) the parameter exceeds a predefined limit, which could cause damage
to the equipment if operation continues under such a condition, and (2) the anomaly pattern deviates from a normal
operational pattern even below the predefined threshold, which could be a symptom of potential failure or improper
operation. Therefore, early detection of these anomalies is critical in proactively avoiding forced outages or part damage,
thus reducing property loss and reducing maintenance costs.
At GE, there is a multi-generational approach to anomaly detection using a combination of physics-based knowledge, fleet
knowledge and collected sensor time series data. Each asset has algorithms that draw from both physics and data, with
models constantly evolving as assets move through their life cycle. The five types of technologies employed at GE include:
Domain or physics-based methods. Here, Digital Twin physics based models enable anomaly detection within the
physical plant through the comparison of calculated parameters with measured values. By analyzing the expected values
throughout the plant with the reported values, GE can determine whether a true physical anomaly is occurring or if a sensor
out of calibration. If a sensor issue is identified the physics based models can be used to provide a “virtual” sensor reading
based on the remaining valid data. If a true anomaly is detected, then the Digital Twin applications can provide action
recommendations to minimize potential impacts.
Statistical process control. Univariate and multivariate control chart techniques are used with thresholds set empirically or
from domain expertise. These include algorithms to detect operational regimes so the appropriate filters and control chart
limits are applied (ex: to account for steady-state base and part load operations, as well as transients).
Machine Learning anomaly detectors. AT GE, a full suite of machine learning algorithms are used, ranging from
multivariate multi-level survival models to baseline asset risk, to classification techniques like logistic regression, decision
trees, random forest methods, neural networks and clustering methodologies. These models are usually derived using
healthy and fault data using GE’s historical database of sensor and configuration data.
Advanced signal processing techniques. There are certain failure modes, especially in gas turbine combustion systems
and compressors, where advanced signal processing techniques are needed to detect subtle anomalies in the presence of
sensor noise. A variety of algorithms that use wavelets, kernel regression and multi sensor data fusion techniques are used
as needed.
Deep Learning Anomaly Detection. GE has invested in cutting-edge artificial intelligence technologies like deep learning
neural networks to detect anomalies. GE leads the industry in the application of these methods for anomaly detection for
operational sensor time series data.
We measure the impact of anomaly detection methods using metrics like failure mode coverage, probability of fault
detection, accuracy of fault diagnosis, false alarm rate, early warning time, etc. These metrics are evaluated frequently by
GE’s analytic staff and if needed, models are tuned, modified or redeveloped as emerging faults are detected throughout the
life of the asset.