PEMS Examples
Introduction:	Because of reliability and cost issues, the U.S. Environmental Protection Agency (USEPA) has for some time accepted the replacement of Continuous Emissions Monitor Systems (CEMS) by PEMS. PEMS are “virtual sensors”, which use mathematical models to predict values of variables that are difficult or expensive to measure. The models are developed empirically from sample data, and correlate the predicted variable to other process variables that are readily measured. Empirical methods include statistical regression, polynomials, and artificial neural networks (ANNs)[1]. PEMS can be used as a replacement for CEMS[2] and/or to provide a process/emissions model for advanced control. They are developed using data collected by calibrated CEMS, which are optionally removed when the PEMS becomes operational[3]. Two projects for predicting various chemical stack emissions are described below.
WFI Power Boiler Example:	Power boilers are fairly simple processes in which fuel is burned to generate steam. Pulp and paper manufacturing is highly energy intensive, and power boilers are used for many purposes. The commodity nature of this industry forces producers to be extremely efficient, and the boilers used are often capable of using many types of fuels to minimize costs, including waste products (bark and black liquor), coal, oil, and gas. The PEMS project described here targeted a type of boiler called a waste fuel incinerator (WFI), which burned bark, coal, oil, and gas. Emissions were known to vary considerably with fuel types. The project’s purpose was to create a PEMS for advanced, model-based control, rather than replace the CEMS already installed. The emission of primary interest was nitrogen oxide compounds (NO), however, sulfur dioxide (SO), free oxygen (O), and opacity were also evaluated. A existing plant data historian recorded measurements for these emissions and other process variables, including: Ambient Air Temperature Amount Bark Fed to the Boiler Air Pressure Over and Under the Bark Bed Amount Coal Fed to the Lower, Middle and Upper levels Burner Levels 1, 2 and 3 Flame/Gas Temperature Amount Natural Gas Fed to the Igniter and Main Burners Amount Fuel Oil Fed to the Lower and Upper Levels Primary and Secondary Air Feed to the Lower, Middle, Upper Levels Output Steam Pressure, Flow, and Temperature For the PEMS to predict emissions over a wide range of operating conditions, it was necessary to vary fuel types and set points for air sources, temperatures, steam flow and pressures[4]. The stack emissions were measured using the existing CEMS, which were calibrated one week before data acquisition. The data was collected over a two-week period, and comprised 1,340 15-minute averages. Modeling results were as follows. NO_x (Measured/Predicted): R² = 0.95 and RMSE[5] = 7.7 ppm, accurate enough for RATA certification. By senstivity analysis[6], NO_x was found to be most dependent on Output Steam Flow, Ambient Air Temperature, Air Pressure Over the Bark Bed, Amount of Coal Fed to the Upper Level, and Burner Level 1 Flame/Gas Temperature. SO₂ (Measured/Predicted): R² = 0.93 and RMSE = 11 ppm. SO₂ was found to be most dependent on Output Steam Temperature, Air Pressure Under the Bark Bed, Natural Gas Fed to the Main Murners, Air Pressure Over the Bark Bed, and Burner Level 2 Flame/Gas Temperature. While the statistical accuracy of this model was good, it was seen to miss periodic spikes (green arrows) due to a cause that was later traced to daily operator maintenance of a filter. Once accommodated, the model would be acceptable for RATA certification. O₂ (Measured/Predicted): R² = 0.97 and RMSE = 0.44%, accurate enough for RATA certification. O₂ was found to be most dependent on Output Steam Flow, Air Pressure Over the Bark Bed, Air Pressure Under the Bark Bed, Amount of Coal Fed to the Upper Level, and Burner Level 2 Flame/Gas Temperature. OPACITY (Measured/Predicted): R² = 0.73 and RMSE = 0.30%. Opacity was found to be most dependent on Output Steam Temperature, Air Pressure Under the Bark Bed, Burner Level 3 Flame/Gas Temperature, Secondary Air Feed to the Middle Level, and Ambient Air Temperature. In comparison to the NO_X, O₂, and SO₂ models, the statistical accuracy of this model was poor and insufficient for RATA certification. One can see evidence (dashed green line) of one or more “unmeasured disturbance variables”, whose long term behaviors are not represented in the candidate input variables. Weather variables such as rainfall, barometric pressure, dew point, and humidity are often the culprits.
Lime Kiln Example:	Large kilns, fired by a variety of fuels, are used to reclaim lime from pulping process waste streams. The PEMS project described here focused on predicting particulate matter, but totally reduced sulfur compounds (TRS) was also evaluated. Interviews with several process experts and kiln operators identified 60 parameters as potential inputs to a model, with a subset expected to be the most likely contributors to high particulate emissions. A designed experiment was conducted over a one-month period on the target kiln. The data represented 1-minute averages, which were manually edited to remove shutdowns and other segments when instruments or communications failed. The process variables measured included: · Air Temperature, Barometric Pressure, Relative Humidity · Scrubber Bull, Tangential Flows, Recycle Water Flow and pH · Scrubber Dilution Flow, Level, DP · Kiln O₂, Rotation Speed, Hood, Middle, and Front End Temperatures · Mud Density, Flow and Tons · Mud Filter Vacuum, Number of Mud Showers and Shower Temperature · Mud Washer Dilution, Density, Torque and Flow · Oil Flow and Atomizing Steam Pressure · Primary Air Damper Position and Air Flow · ID Fan Damper Position and Speed · Lime Discharge Temperature · White Liquor Cone Dilution, Clarifier Density, Flow, and Torque · Mud Free Lime, Soda, and Solids · Recycle Solids · Residual CaCO₃ and Lime PARTICULATE MATTER (Measured/Predicted): R² = 0.91 and RMSE = 2.5 gr/ft³, accurate enough for RATA certification. Particulate matter was found to be most dependent on Kiln Front End Temperature, Oil Flow, Atomizing Steam Pressure, Kiln Middle Temperature, and Recycle Water pH. TRS (Measured/Predicted): R² = 0.87 and RMSE = 0.64 ppm. These results made possible RATA certification questionable without additional work.. TRS was found to be most dependent on Mud Density, Kiln Middle and Front End Temperatures, and Mud Shower Temperature.
Conclusions:	Based the above examples, the following conclusions can be drawn about the utility and development of PEMS. · PEMS are a desirable alternative to CEMS when: - CEMS installation and operating costs are extremely high. - Hardware sensors are untenable for extreme applications, such as wet stack opacity and particulate matter. - A process model with “what ifs” capability is needed for advanced control and process optimization. · PEMS can be applied to a wide range of emissions types, including NOX, O2, SO2, opacity, particulate matter, and TRS. · PEMS accuracy depends on having data that covers the entire operating range of interest, which can include routine maintenance activities. The data may be collected from: - Long term historical data archives. - A designed experiment over a shorter period. · With good data in hand, PEMS model development is generally straightforward. · PEMS deployment is also generally straightforward, and can use existing sensors, data acquisition/archival, and SCADA[7] or DCS[8] infrastructure. [1] ANNs are a multivariate, non-linear modeling technique developed from the field of machine learning, a branch of artificial intelligence (AI). For complex problems that are well characterized by data, they are often superior to statistical and polynomial models. [2] USEPA requires CEMS to be certified for accuracy once a year using a process called a RATA test (relative accuracy test audit), which requires evaluating performance over a wide range of operating conditions. PEMS are subjected to the same test and performance criteria when they replace CEMS. This is achieved by temporarily installing CEMS to which PEMS are compared during RATA tests. [3] Models are developed using actual data for calibration (“training data”). In order to assess their accuracy, data other than that used for calibration must be used for testing (“test data”). The calibration and test data are often taken from the same originating data set by means of various types of filters. [4] Robustness of ANN models can be improved through designed experiments or by collecting a large amount of historical data. [5] RMSE (root mean square error) is an estimate of the predictive error of a model. Unlike the dimensionless R² (coefficient of determination), RMSE has the units of the output variable and is relative to its dynamic range. [6] A sensitivity analysis ranks a model’s input variables according to the impact they have on the output. It is performed by systematically ranging each input and computing how much the output changes. [7] Supervisory Control and Data Acquisition System - used by many manufacturers for process control. [8] Distributed Control System - a process control technology that is generally more capable and expensive than SCADA.