Data mining

There are many ways to reduce energy consumption and to improve operational efficiency. Good energy management practice can reduce the costs of utilities, but once good practice is in place, significant additional savings can be realised by tackling the manufacturing process itself. Often all that is required is to ensure that the existing control systems are working and are properly tuned.

Sometimes, however, the nature of the process means that more is required. Modern Scada/PLC systems allow plant operators routinely to amass enormous amounts of operating data. Within these data is information knowledge about the process behaviour that could help identify opportunities for process improvements. Extracting this information knowledge by data mining can identify hidden performance patterns, turning process operational data into useful knowledge.

Data mining techniques have been widely applied in the financial, insurance and marketing industries to extract valuable information from masses of data. These ideas are now being applied in the process industries. Cleveland Potash, for example, has used data mining to improve rotary kiln dryer control at its Boulby Mine. With direct support from the Carbon Trust's Action Energy programme (formerly the Energy Efficiency Best Practice Programme), the company achieved savings of £47 000 per year with a payback time of only five months.

Cleveland's operations at Boulby are complex and energy intensive, with an annual energy bill of £6.8million. An initial study to determine what might benefit from a data mining approach, included underground operations and, above ground, flotation, drying, compaction and milling. It was eventually decided to concentrate on a rotary kiln dryer because previous studies had indicated that there was significant potential for fuel savings, extended plant life, reduced emissions, improved reliability and a more consistent product quality. And operating data was already available.

The dryer processes 60tonne/hr of wet product from flotation cells. The purpose is to vaporise all the water, which varies with feed consistency and centrifuge efficiency. These are large units; the shell is 3m in diameter and 20m long. It rotates at 8rpm. At the front end is a refractory lined combustion chamber with a fuel oil burner. A forced draught fan supplies air. Dampers in the air duct are linked to the fuel valve to control the air-to-fuel ratio.

The exhaust gas temperature is maintained automatically by adjusting the fuel flow controller. This setpoint is manually adjusted to achieve an adequate product temperature.

Examination of plant data, collected over a 4-month period covering 900 hours of typical operation, revealed the following:

(a)there were variations in product temperature implying excess energy consumption.

(b)the gas temperature controller was poorly tuned.

(c) the product exit temperature could be improved.

(d) the average product temperature could probably be substantially reduced.

Decision trees (see sidebar below) were generated and used outcomes such as oil firing rate, product and exhaust gas temperatures. Attributes were measures such as centrifuge current, indicative of product load and moisture content.

Some of the results and associations were obvious but others were subtle indicating, for example, the effect of centrifuge operation on dryer stability.

It was clear that improved product exit temperature control would lead to more energy-efficient operation. A predictive model for the product temperature using rule induction was developed and used as the input to a PID controller, which provides a setpoint trim for the existing exhaust gas temperature controller. The new system can be easily adapted to accommodate additional knowledge in the future and it is easy to revert to the existing (less efficient) strategy in the event of a failure of the advanced control system - a necessary confidence-building feature.

The resulting fuel saving of £47,000 per year stems directly from the improved control of product temperature. This originally operated at a mean of 216 degrees C +/- 30 degrees C, but the new system achieves a mean of 185 degrees C ±5 degrees C. The study has also led to the identification of further energy saving opportunities. For example, using variable speed drives to improve the dryer draught control will save a further £24,000 per year.

The dryer data mining study took 30 man-days and cost £30,000 with £5000 for hardware and software. The fuel saving on the dryer alone gave a payback period of just five months.

Sidebar: Mining out the information nuggets

The main technique used in data mining to improve process control is 'rule induction'. This helps to identify patterns and to express them as rules, which are expressed graphically as a decision tree.

Rule induction is based on splitting up the data set using influencing factors (attributes) to reach an outcome. An outcome might be, for example, a temperature setpoint, while an attribute might be the moisture content of a feedstock. The route to the outcome describes the criteria for that rule.

Rule induction has particular advantages over some other 'black box' modelling methods - models can be implemented very quickly from large data sets, and they are 'transparent' enabling the validity of each individual rule to be established.A data mining study, such as that at Cleveland Potash, essentially involves the following:

1. Acquire relevant data from the monitoring and control system. This is not trivial. It is critical that sufficient data from the process are gathered.

2. Pre-process the data to remove obviouslyirrelevant information such as periods of plantdowntime, but also the non-obvious. Pre-processing is an important and time-consumingtask if carried out properly. The success of theproject depends upon it.

3. Analyse the data. This should use both conventional as well as data mining techniques.Don't forget your engineering judgement.

4. Develop the decision tree by rule induction.

5. Implement the improved operational procedures or control system based on the results.

Professor Jim Anderson is with Anderson Barr Consulting and Dr. Alan McCullough is with Action Energy.