Using Data Mining Methods to Optimize Boiler Performance: Successes and Lessons Learned Thomas Hill, Ph.D. StatSoft Power Solutions StatSoft Inc. www.statsoft.com www.statsoftpower.com Optimize combustion, OFA l Stabilize operations l Reduce emissions l Predict problems Headquarters: StatSoft, Inc. 2300 E 14 St Tulsa, OK 74104 USA +1(918) 749-1119 Fax: (918) 749-2217 info@statsoft.com www.statsoft.com Australia: StatSoft Pacific Pty Ltd. Brazil: StatSoft South America Ltda. Bulgaria: StatSoft Bulgaria Ltd. Chile: StatSoft South America Ltda. China: StatSoft China Czech Rep.: StatSoft Czech Rep. s.r.o. Egypt: StatSoft Middle East France: StatSoft France Germany: StatSoft GmbH Hungary: StatSoft Hungary Ltd. India: StatSoft India Pvt. Ltd. Israel: StatSoft Israel Ltd. Italy: StatSoft Italia srl Japan: StatSoft Japan Inc. Netherlands: StatSoft Benelux Norway: StatSoft Norway AS Poland: StatSoft Polska Sp. z o.o. Portugal: StatSoft Iberica Lda Russia: StatSoft Russia S. Africa: StatSoft S. Africa (Pty) Ltd. Spain: StatSoft Iberica Lda Sweden: StatSoft Scandinavia AB Taiwan: StatSoft Taiwan United Kingdom: StatSoft Ltd. USA: StatSoft, Inc. Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc.
Overview Data mining methods have proven increasingly useful for untangling complex relationships in high-dimensional process data (with hundreds or even thousands of parameters) These methods can produce significant performance improvements from existing equipment and without requiring any additional hardware or equipment modifications This presentation: Discusses how data mining methods and algorithms are different from simple trending and/or the application of traditional statistical methods Presents an overview of typical workflows and necessary steps from data preparation through modeling and optimization Reviews typical examples and case studies where data mining methods have been applied successfully for combustion optimization Summarizes specific lessons-learned ( how to be successful ) Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc. 2
Data Driven Technologies Historical operational data can be used for process optimization, by applying methods that: Focus on knowledge discovery, detection of patterns, clusters, etc. Apply advanced knowledge discovery (data mining) algorithms that will result in valid models for highly complex data E.g.: recursive partitioning, stochastic gradient boosting/bagging of trees, multivariate adaptive regression splines, support vector machines,... All of these algorithms are universal approximators, which can model (approximate) any relationships between parameters (non/linear, interactions, etc.) Use optimization methods with the goal to achieve robust/stable and optimized performance (e.g., low-variability NOx and CO, in presence of normal variability in fuel quality, load, etc.) Carefully control, during optimization, the extent to which models interpolate/extrapolate from actual data (empirical evidence) Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc. 3
Data Driven Technologies: Comparison to Other Methods Unlike CFD (Computational Fluid Dynamics) modeling We analyze the actual data describing actual observed performance of the power plant over the past 1 or 2 years, or more Unlike DOE (Design of Experiments) methods We do not perform trial-and-error testing to find simple relationships, but identify complex relationships through the application of nonlinear general modeling algorithms ( data mining ) To use a metaphor: The power plant or boiler is talking to the operators or control system through the language of numbers, recorded into a data historian such as OSI Pi The data mining methods and algorithms described on the previous slide decipher this language, and leverage the extracted information for process optimization Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc. 4
Data Mining and Statistical Modeling Knowledge Discovery vs. Statistical Analysis Statistical Analysis Focuses on hypothesis testing and parameter estimation Fits parsimonious statistical models with the goal to explain Examples: Regression, traditional generalized linear models (GLM) Data Mining and General Predictive Analytics The data are your model! Focuses on the detection of repeatable patterns in historical data, with the goal to make predictions about future events Focuses on knowledge discovery, detection of patterns, clusters, and so on; we only have data and no (or few) expectations and hypotheses Applies pattern recognition algorithms or general approximators, with goal to capture valid/repeatable relationships among variables in the data Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc. 5
Data Driven Technologies: Summary In short, based on these methods, StatSoft developed and validated a workflow that: Uses all historical data that are already routinely collected (and have already been collected for some time) Identifies in those data the specific operational parameters that are critical for optimal boiler performance Builds data mining models describing how exactly the important parameters affect the performance of a furnace Uses those models for robust optimization (optimization for robust, low-variability operations/performance) Identifies optimized parameter ranges and relationships for critical operational parameters that can be implemented into the existing control system Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc. 6
Concrete Examples: Cyclone Furnaces See also: EPRI/StatSoft Project 44771: Statistical Use of Existing DCS Data for Process Optimization (2008) Numerous projects have applied data mining technologies to cyclone furnaces of various designs and sizes Goal typically is to stabilize flame temperatures within a desirable range (e.g., measured flame intensities, door, slag temperatures..) Robust high flame temperatures are usually related to reduced emissions Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc. 7
Concrete examples: Cyclone Furnaces (cont.) Typical results are often immediate Significant sustained improvements to LOI and boiler efficiency (>20%) without negative side-effects have also been demonstrated Quote:.. after the data mined settings were put into the control system (unit 3) with solid success ( $1.4M savings in oil), however the results were harder to quantify in unit 4, as 4 never needed or used as much oil as unit 3. (Oil in unit 3 was more of a crutch than on unit 4.) However, the settings (unit 4) that you folks did suggest, did lower the NOx across the entire load range. Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc. 8
Concrete Examples: Cyclone Furnaces (cont.) After implementing results into the DCS system as new defaults, flame temperatures are consistently higher Nox and CO emissions are significantly lower Conclusion: The new settings worked as expected and were highly effective at lowering emissions Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc. 9
EPRI Perspective See: http://my.epri.com/portal/server.pt?abstract_id=000000000001016494 Results The results show that the data analysis and optimization methods effectively identified specific ranges for a relatively small subset of operational parameters. Significantly improved and stable operations resulted for all cyclones. Application, Value and Use These methods provide a way to achieve cost-effective and realistic (virtually immediately obtainable) boiler optimization given the existing data and control systems without the need for further boiler modifications or hardware and/or software purchases requiring only modifications to the parameters and equations guiding existing control system software. Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc. 10
Concrete Examples: Wall Fired Several projects have focused on lowering NOx and CO emissions (or SCR inlet NOx) over the entire load range. For example, optimization of a 400 MW Coal-Fired DRB-4Z Burner for consistent/robust low-nox operations under low load (50-175 MW) Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc. 11
Concrete Examples: Wall Fired (cont.) Typical Results: Lower NOx during testing and more robust performance (lower variability in NOx measurements, with fewer/no spikes) Continued improvements after formal validation testing ended Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc. 12
Concrete Examples: T-Fired To date StatSoft has completed only one project with a twin T-fired boiler, with a complex (multi-zone) OFA system Goal was to stabilize and reduce CO while maintaining or lowering Nox Simple before-after tests showed the effectiveness of specific combinations of settings for various parameters (involving Windbox/Furnace Differential Pressure, and OFA parameters) Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc. 13
Lessons Learned (1) The methods described here are generally applicable to any type of coal furnace And likely other types of furnaces and complex equipment as well Only basic requirement is a process historian (database, such as OSI Pi) that collects and stores operational parameter data The methods described here can be used to optimize boiler performance using the existing control systems and methods Lessons learned (organizational): Commitment by operators and performance engineers to project is critical for success Identify at the start how the results are to be implemented (e.g., into the control system); this decision will guide the modeling efforts and ensure that results are actionable Determine the dollar value of the performance improvements; this is critical for buy-in by all stake holders Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc. 14
Lessons Learned (2) Lessons learned (data and methods): Data quality must be ascertained before performing any analyses; in particular: Make sure that data gathering devices, sensors, scaling, or computations did not change Establish through testing that the variability of key inputs (e.g., secondary air flows) are controllable within the required bounds; in particular, if the unit and air flows swing (e.g., because of unreliable O2 probes/measurements), then recommended settings may not be achievable! Sometimes, changes to the logic of the control system are necessary Examine causes of variability in inputs (e.g., air flows), to determine that operators do experiment and drive the unit ; sometimes, all variability is produced through default operation of the control system It may be necessary to introduce variability to the data through systematic testing, DOE Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc. 15
References Nisbet, R., Elder, J., & Miner, G. (2009). Handbook of Statistical Analysis and Data Mining Applications. Academic Press StatSoft Electronic Text Book: www.statsoft.com/textbook/stathome.html Also available from Amazon: Hill, T. & Lewicki, P. (2005). Statistics: Methods and Applications. EPRI/StatSoft Project 44771. Statistical Use of Existing DCS Data for Process Optimization. EPRI, Palo Alto, CA: 2008. Copyright StatSoft, Inc., 1984-2012. StatSoft, StatSoft Power Solutions, StatSoft logo, and STATISTICA are trademarks of StatSoft, Inc. 16