The power generation sector has been challenged by trends in digitalization, growth of intermittent renewables, equipment performance improvements, new generation, and storage technologies. Digitalization is a broad concept that can be used as a tool to achieve multiple goals in this new scenario where performance and availability are critical.
Continuous optimization during the operation cycle of assets can be difficult. Even so, the growing capacity of computing processing enables plant owner-operators to achieve profit margins even at very high thresholds. As systems become more complex, human analysis and decision-making can be augmented with algorithms and software tools that standardize operational decisions and help achieve maximum performance. Decision support systems can go beyond human capacity because they can factor in and correlate multiple complex performance variables and operator experiences, delivering the most desirable outcome.
Operators are responsible not just for maximizing performance, but also for safety, asset management, and other technical and strategical aspects. Considering these multiple goals, advisory systems are tools that help optimize plants but do not automate the decision process, avoiding a failed decision that is not modeled. An advisor system displays suggested best conditions, and then the operator can decide whether or not to proceed. This approach enables long-term reliability testing and a modeling certification process aimed at checking if future decisions can be made automatically.
Optimization modeling is the main challenge when developing an advisor system. Modeling can be diverse, involving mathematical and/or data models. To explore different modeling and evaluate approaches, this article presents four optimization solutions, three on thermal power plants and a fourth on a cascade hydropower complex.
Multi-Level Energy Efficiency Monitoring
Energy performance monitoring is the first step in developing an optimization strategy. A monitoring tool provides optimization when it enables the identification of gaps and delivers insights from operational comparisons. There are two main examples of deviations: design versus actual conditions, and real-time versus historical or forecast parameters.
When showing good correlations with existing data, first principle equations are very useful tools to model the performance of operating units. To develop this real-time efficiency platform, Radix Engineering and Software used well-known thermodynamic equations combined with real-time data coming from distributed control systems. This approach was developed and tested in the power generation park of Petrobras, the Brazilian national oil, gas, and energy company. The Petrobras system is composed of various power plant sizes (from 87 MW to 1,058 MW) and configurations—natural gas (simple, combined cycle, and cogeneration) and diesel oil units—totaling more than 5.7 GW of capacity.
The platform is capable of monitoring the energy efficiency of the thermoelectric units in real-time and inferring the expected efficiency, allowing the rapid identification of operational deviations. Prior to implementation, the power plants had only targets for energy and steam dispatch, without direct monitoring of efficiency. For this reason, operators had great difficulty identifying efficiency deviations.
Efficiency modeling considers three hierarchical levels in the production system: the plant as a whole, each process unit, and the main equipment. For example, inside a plant, for each of the specific processes of steam generation, water treatment, compressed air, refrigeration, cooling towers, natural gas compression, and others, there is a specific thermodynamic model to enable the energy efficiency calculation. In the same way, there are specific models for monitoring the energy performance of each main equipment, such as gas turbines, combustion engines, heaters, boilers, compressors, and pumps. Furthermore, two complementary methodologies were tested to design a forecasting model: multivariable regression and natural networks. Both presented equivalent good results.
To this end, it is important that the relevant process variables are available and reliable. To address this challenge, a detailed diagnosis of the available instrumentation was performed to identify gaps in the monitoring system, in order to enable effective energy management of the plants.
A Decision Tree to Optimize Decision-Making
The University of Massachusetts, Amherst has one combined heat and power (CHP) plant to supply steam and energy to meet campus demands, relying on several different types of equipment like gas and steam turbines, boilers, photovoltaics, and batteries as a supplementary source when needed. Annual steam and power demands vary widely. Steam consumption is higher during cold seasons and lower during hot seasons. Power demand has an inverse behavior, with peak consumption during hot seasons. The variability of the operations (including power and steam generation, selection between three diverse burner fuels, the optional utilization of chillers and boilers, and irradiance availability) creates a complex environment for decision-making.
The first step to support the operations team was the consolidation of the multiple data sources including plant equipment, power meters, field flow meters, and fuel price information. These data allow the calculation of the main key performance indicators (KPIs) related to the operating costs and energy efficiency of the plant.
However, because the CHP plant can be operated in multiple ways from the Energy Command Center, displaying the data was not enough. The second step was to create a decision tree, with a rules engine based on results of offline thermodynamic simulations, to achieve the most efficient operating conditions for the plant. The aim of this application was to facilitate remote operations monitoring and a straightforward decision-making process for engineers and supervisors.
The methodology developed by Radix combines an engineering assessment, based on the simulations, with the development of an operational intelligence system. The advisory system helps the plant operate at its maximum possible efficiency. The optimization model was the main challenge of the system. It required combining an approach that could indicate the best operational configuration considering multiple scenarios and running in real-time a detailed model with all main equipment linked.
1. University of Massachusetts, Amherst Energy Command Center decision tree advisor screen. Courtesy: Radix
The solution is an online decision tree (Figure 1) that stores 150 scenarios simulated offline under distinct configurations. During real-time operation, the application analyses the online variables and indicates the best configuration to operate. The system consolidates several data sources including plant equipment, power meters, field flow instruments, and real-time and day-ahead fuel price information (including ultralow-sulfur diesel, natural gas, and liquified natural gas). This set of data feeds the optimization and efficiency models, forecasts, and KPIs that will be displayed and accessed in the system.
The main results include the reduction of operation cost by 3% or approximately $900,000 per year by making real-time decisions on major equipment alignment based on current energy market prices.
Data Modeling and Artificial Intelligence for O&M Optimization
When thermoelectric plants are used as a reserve source of energy, they are paid for availability. In these cases, the optimization of availability, reliability, and performance of the generating machines are critical issues to maximize the economic results of the business and to guarantee the supply to the electric sector. To overcome the challenge of improving the availability, Radix is developing two projects using data modeling to optimize operation and maintenance (O&M) together.
The processing and analysis of large amounts of data in real time led to the appearance of the Industry 4.0 nomenclature, which is supported by concepts such as machine learning, business intelligence, big data, Internet of Things (IoT), cloud computing, among others. Currently, the Radix research and development (R&D) team is developing projects focused on the application of predictive maintenance (PdM) and energy optimization, using data science and artificial intelligence, for thermoelectric plants.
The benefits and results of these projects are numerous and involve the predictive identification of trends, deviations, failures, and critical inefficiencies. This increases the useful life of components and equipment, and consequently reduces maintenance costs. It also increases energy efficiency, and therefore reduces energy costs, greenhouse gas emission rates, and environmental impact.
The fact that PdM and optimization projects are executed together is a differentiation, because one system will serve as an input for the second. While many practices adopt the engine health index in a static way, these projects use them dynamically, assessing the health of the generating units through a digital twin and deep-learning approach. This work is been developed under the R&D program of ANEEL (the Brazilian Electricity Regulatory Agency) financed by Epasa—Centrais Elétricas da Paraíba S.A.
Cascade Hydropower Optimization
Brazil has a predominantly hydropower generation matrix and the coordinated operation of plants is decisive for the interconnected electrical grid. The hydroelectric scheduling program is organized in long- to short-term horizons. Long-term analysis makes use of stochastic inflow data and simplified equivalent units. Short-term analysis, on the other hand, makes use of deterministic inflow data, and detailed information of power plants and reservoirs. These challenges are more complex when it concerns cascade hydropower plants with small reservoirs.
An R&D project was carried out to develop a solution to address the optimization of a complex of three cascade hydropower plants (330 MW total) belonging to the same owner-operator. The system focuses on the short-term optimization to help operators make real-time decisions, including daily scheduling. The decision variables are the state of the generator and the power dispatched by each generator in operation. The complexity and the dimensions of the problem grow as the number of intervals or units considered increases, becoming unpractical to most conventional optimization methods.
The mathematical solution considers the hydropower plant model (power produced by a hydroelectric generator, hydraulic head, upstream water level, downstream water height, efficiency of the hydro turbines—represented by a function known as the turbine Hill Chart—and head losses area) plus reservoir equations. Three different metaheuristics were tested: Particle Swarm Optimization (PSO), Genetic Algorithms (GA), and Differential Evolution (DE). The DE method produced the best results considering the best cost-benefit relation for a minimum objective function cost and a rapid convergence.
The system has four modules—daily scheduling, monitoring, real-time optimization, and indicators—that together integrate all of the utility’s planning and operations in an optimized manner. During operation, a 1% average improvement in the efficiency level of the complex has been measured, from approximately 92% to as high as 94% after system deployment. Future developments in this solution will consider data modeling to tune best parameters of the optimization model.
This work was developed under an R&D program of ANEEL (the Brazilian Electricity Regulatory Agency) as part of the Cascade Hydropower Plants Optimization Project financed by CERAN (CIA Energética Rio das Antas).
Optimization Models Work
The solutions presented show how it is possible to use different modeling techniques to optimize power plants focused on performance and availability improvements. As power plant processes and equipment are based on consolidated engineering knowledge, optimization by first-principle equations modeling can be a first approach to consider in a variety of solving methods and simulations tools. Furthermore, data-modeling approaches are also useful and can be used as the main model, as well as to improve the calibration of first-principle and optimization models. ■
—Tassio Simioni is energy business manager, Priscila Vieira Gameiro is a control and automation engineer, Flavio Leite Loução Jr. is a project coordinator with Radix Engineering and Software (radixeng.com).