Operating system description

Off-take from these tanks can proceed either to one of two pipeline tanks or to one of two delivery tanks, or for direct supply to the sea or road. The pipeline tanks provide stock for pipeline delivery, whilst the delivery tanks provide stock for dispatch by road or by sea.

A key feature of tank operations is the fact that at any one time a storage tank is either being emptied, filled or held at a constant level. Tank level measurements in Figures 3.2 and 3.3 show that these stages seldom coincide. The x-axis in these figures represents the hourly sample points when measurements were taken.

Coincidence of filling, emptying or constant stages would be presented by a low gradient increase or decrease in the tank levels. This phenomenon would be due to a simultaneous use of pumps in front of and behind a tank. In this case the pump capacities on the two sides of the tank were not equal therefore due to a bigger filling pump capacity, a slow but constant inflow could be examined.

Figure 3.1 Tank network of ULG95 at Stanlow

IN-LINE BLENDER

Blender Tank1 Blender Tank2 Blender Tank3 Blender Tank4

Pipeline Tank1 Pipeline Tank2 Delivery Tank1 Delivery Tank2

Furthermore, tanks are filled or emptied at a constant rate. Between filling and emptying, tanks are normally held at a constant level for quality control reasons. If the tank level reaches a certain low level between the filling and emptying stages, the quality of the crude oil is tested to avoid the presence of unwanted chemical properties. On the other hand, the tank level is not allowed to go above a certain level for safety and fluid dynamics reasons.

The “flat” periods of the stages where the tank is kept at a constant level can be seen in Figures 3.2 and 3.3. In Figure 3.2 the time span of the flat intervals varies but in general flat periods occur more often at low levels than after filling. The length of the constant stage depends on the actual pipeline or road delivery demand pattern for the gasoline. Another boundary condition, which is applicable to the analysis, is that tanks are always filled and emptied in sequence. Only rarely are two tanks emptied or filled simultaneously.

Figure 3.2 Four blender tank level time series

0 500 1000 1500 2000 2500 3000 3500 4000

0 1

2x 10⁴ Four blender tank time series

Stock (m3 )

0 500 1000 1500 2000 2500 3000 3500 4000

0 1 2x 10⁴

Stock (m3 )

0 500 1000 1500 2000 2500 3000 3500 4000

0 1 2x 10⁴

Stock (m3 )

0 500 1000 1500 2000 2500 3000 3500 4000

0 1 2x 10⁴

Stock (m3 )

Red: Pipeline 1 Blue: Pipeline 2 Green: Delivery 1 Black: Delivery 2

In Figure 3.3 the red and blue plots show the time series for the pipeline tanks, the green and black plot exhibit the time series for the two delivery tanks. The time interval of the flat stages of the delivery tanks is shorter, this is due to the nature of the transportation arrangements of tank cars. The variability of a process can be classified into two general types (Montgomery, 1991). There are many small, essentially unavoidable sources of variability that are inherent to the system itself. These are termed chance variation or “common case variation”. Chance variation is predictable over time due to its randomness, but, it cannot be easily reduced or eliminated from the process.

Examples of chance variation in the process include variation due to raw materials or thermal and electrochemical noise. A process that exhibits only chance variation is said to be in a state of statistical control.

In addition, there are other sources of variation that are only occasionally present in the process. This type of variation forces an otherwise stable process to become unstable and unpredictable. Furthermore, it represents an unacceptable level of process performance and is termed “special cause variation” or “assignable variation” due to the fact that it can readily be assigned to particular causes. Although, assignable cause variation is relatively large when compared to chance variation and is not predictable over time, it can typically be mitigated by applying appropriate corrective actions to the process. Example sources of assignable variation of the process included different flow meter set-up conditions, different suppliers of raw materials, joining of different sub-supply pipelines. When the process exhibits only assignable cause variation, it is said to be out-of-statistical-control.

Figure 3.3 Two pipeline and two delivery tank level time series

0 500 1000 1500 2000 2500 3000 3500 4000

0 5000 10000

Two pipeline tank and two delivery tank time series

Stock (m3 )

0 500 1000 1500 2000 2500 3000 3500 4000

0 5000 10000

Stock (m3 )

0 500 1000 1500 2000 2500 3000 3500 4000

0 5000 10000

Stock (m3 )

0 500 1000 1500 2000 2500 3000 3500 4000

0 5000 10000

Stock (m3 )

Hourly tank level measurements showed a significant number of so-called, "non-physical values", which are caused by the relatively large variation of the tank level time series data. These values are referred to as "spikes". These spikes must be dealt with prior to undertaking a subsequent analysis. The spikes were removed by applying a sequence of corrective actions, not detailed here. Without the corrective actions false conclusions would be drawn at the data analysis stage. Further examples of the original time series, are given in Appendix IV.

Exploratory data analysis is more than simply providing a summary of data measures and various graphical displays, it includes checking that the data are consistent with expectations (Kresta, et al, 1991) . The physically acceptable values of filling and emptying rates fall into pre-defined regions since the filling and emptying capacity of these tanks are finite and limited by the diameter of the supplying pipelines. Since the data has already been collected, the first stage of the analysis was to examine the data - data pre-screening. Pre-screening focuses on the visual screening of the data. No mathematical manipulation is used at this stage. Through initial plots of the data, time series plots, histograms and scatter plots a feel for the data can be obtained, any non-confirming observations identified subjectively, natural clustering of the observations detected and the underlying distributional properties examined. Appendix IV shows the important properties of the original data.