INDUSTRIAL ENERGY MANAGEMENT

CUSUM - a monitoring and targeting technique

(Incorporating distance learning material of D R Croft)
CUSUM stands for CUMulative SUM deviation. It is a statistical technique which can be
used as a management tool for a wide range of processes, but particularly energy
consumption.
Peter Harris in his book on CUSUM1 states that it was first developed in the early 60s, but
in essence it has been used earlier as regression analysis. During the 2nd World War the
Fuel Efficiency Committee considered a wide range of techniques for energy saving and
regression analysis was one of them. With the advent of electronic calculators, computers
and spreadsheets the technique of CUSUM can now be easily applied. CUSUM is an
extension of regression analysis in that regression analysis predicts a deviation and
CUSUM analyses the deviations.
In management it is necessary to set targets and in Energy Management targets are set for
energy consumption against which the performance of a given process, piece of plant or
building can be measured. If a manufacturing company implements energy-saving
measures then it would expect the energy content per unit output to decrease. There are
circumstances when the simple measure Energy Content per Unit Output does not give a
true or easily identifiable indicator of increased energy efficiency. Consider the
information revealed in Figure 1 which shows the energy content per tonne of production
from a gas-fired furnace over a 30-week period.
THERMS PER TONNE
45
40
35
30
25
20
15
10
5
0
1

11

13

15

17

19

21

23

25

27

29

WEEK NUMBER

Figure 1. Variation in therms2 per tonne

If you were asked the following questions could you answer them from the values shown in
Figure 3:
1
2

"Energy Monitoring & Target setting using CUSUM" Cheriton Technology Publications 1989
Therms were commonly used to measure heat content of fuels = 100,000 Btu = 94,780 kJ

1. How many energy-saving measures have been implemented, and at which weeks?
2. How much energy has been saved by each measure ?
3. Are all the measures still, working ?
Clearly, you cannot answer these questions with certainty; the variation in the value of
(therms per tonne, or specific energy consumption) makes it difficult to be sure what has
happened.
Perhaps the graphs showing the weekly tonnages and therms (Figures 2 and 3) will make
matters easier to understand :
PRODUCTION - TONNES
2000
1800
1600
1400
1200
1000
800
600
400
200
0
1

11

13

15

17

19

21

23

25

27

29

WEEK NUMBER

Figure 2. Variation in production tonnage

THERMS
40000
35000
30000
25000
20000
15000
10000
5000
0
1

11

13

15

17

19

21

23

25

27

29

WEEK NUMBER

Figure 3. Variation in gas consumption

Now it can be seen that when production is high the specific energy use is low, and vice
versa. This suggests that the relationship between energy consumption (E), and production
(P) may be given by:
E = A + (B x P)

where A represents a constant energy use which is independent of production, say the
amount of energy required to keep the furnace on "tick-over", and B is the specific energy
content.
The units of A and B will be therms, and therms per tonne, respectively. It is now clear
that using the indicator of therms per tonne ignores the value of constant A, and that the
error in so doing may be significant at low values of production. The actual numerical
values of A and B have to be determined from the production and therms data of the
furnace. It can be taken that A and B will remain roughly constant for the weeks before
any changes are made and after that, the effect of energy-saving measures should be
revealed in reduced values of, A and/or B. For the case we are considering, we will assume
that it is known that the first energy-saving measure is introduced at the beginning of week
9 (this assumption is discussed later in the section).
Using simple linear regression analysis on the first 8 weeks of data gives the following
formula:
E = 9831 + 15.16 x P
Where E is the energy consumption in therms, and
P is the production in tonnes.
If this formula is correct then there should be little difference between the measured and
predicted values for the first 8 weeks of data. Also the running total (or cumulative sum)
of these differences should be small compared to the values of E. The relevant figures are
shown in Table 1 below with the running total represented by the acronym CUSUM:
Week

1
2
3
4
5
6
7
8

Production
tonnes
(1)
1500
800
600
500
1700
1800
1200
400

Measured
therms
(2)
33000
21500
19500
17000
35000
38000
27000
16500

Predicted
therms
(3)
32570
21958
18926
17410
35601
37117
28022
15895

Difference

CUSUM

(4)

(5)
430
-458
574
-410
-601
883
-1022
605

430
-28
546
135
-466
416
-605
0

Table 1. Comparison of predicted and measured therms for weeks 1 - 8

where columns
column
column
column

(1)
(3)
(4)
(5)

and (2) are the data shown in Figures 2 and 3.

is predicted from the expression E = 9831 + 15.16 x P
= [column (2) - column (3)]
is the running total of column (4)

Repeating the calculation over the remaining weeks should show a more significant
difference between measured and predicted therms if energy saving measures are taking
effect, and hence an ever-changing value of CUSUM. The calculation for the whole 30week period is shown in Table 2:

Week

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

Production
Measured
Predicted
Difference
CUSUM
tonnes
therms
therms
(1)
(2)
(3)
(4)
(5)
1500
33000
32570
430
800
21500
21958
-458
600
19500
18926
574
500
17000
17410
-410
1700
35000
35601
-601
1800
38000
37117
883
1200
27000
28022
-1022
400
16500
15895
605
1100
23900
26506
-2606
900
21100
23474
-2374
400
12000
15895
-3895
500
14600
17410
-2810
1400
27600
31054
-3454
1550
29900
33328
-3428
1650
31300
34844
-3544
1900
34650
38633
-3983
400
16000
15895
105
550
17700
18168
-468
1500
33000
32570
430
800
22000
21958
42
625
15500
19305
-3805
1100
21200
26506
-5306
700
16000
20442
-4442
1700
29000
35601
-6601
1900
34000
38633
-4633
1600
27200
34086
-6886
1200
28000
28022
-22
1900
38500
38633
-133
500
17250
17410
-160
900
23700
23474
226

430
-28
546
135
-466
416
-605
0
-2606
-4980
-8875
-11685
-15139
-18566
-22110
-26093
-25988
-26456
-26026
-25984
-29789
-35095
-39538
-46139
-50772
-57658
-57680
-57813
-57974
-57748

Table 2. Comparison of predicted and measured therms for weeks 1 - 30

A graph of CUSUM against week number is shown in Figure 4. It is now much easier to
see what events have occurred over the 30-week period. The improvements which started
at the end of week 8 continued until the end of week 16 and then stopped being effective.
Further measures were implemented at the end of week 20 which were effective until the
end of week 26. From then on, the furnace performance returned to its initial state.

CUSUM - THERMS (THOUSANDS)

10
0
1

11

13

15

17

19

21

23

25

27

29

27

29

-10
-20
-30
-40
-50
-60
WEEK NUMBER

Figure 4. CUSUM variation over the period

CUSUM - THERMS (THOUSANDS)
10
0
1

11

13

15

17

19

21

23

-10

25

SAVING
~2600 THERMS

-20
-30

SAVING
~3200 THERMS

-40
-50
-60
WEEK NUMBER

Figure 5. Energy savings over the period

CUSUM
10000

DATUM LINE
0
1

9 10 11 12 13 14 15 16 17 18 19 20 21 22

-10000

ACTION 1
SAVINGS

-20000
-30000
-40000
-50000

ACTION 2

-60000
WEEKS

Figure 6. CUSUM for continual savings

The savings accumulated by each action are shown in Figure 6. If we apply regression
analysis to the period of weeks 9 to 16 then the following relation is obtained:
E = 6947 + 14.88 x P
Comparing this to the original expression [E = 9831 + 15.16 x P] shows that the "tickover" energy consumption decreases by a factor
9
8
3
16
9
4
7
2
9
.3
%
9
8
3
1

and the specific energy consumption decreases by a factor

1
5
.
1
61
4
.8
8
1
.8
5
%
1
5
.
1
6
A similar calculation can be carried out on the data for weeks 21 to 26, which will reveal
the expression
E = 5809 + 14.06 x P
The percentage improvements can now be determined.

Concluding remarks
Further information is required from the company to determine precisely what actions
caused the savings to occur and why these actions ceased to take effect. The implication of
the CUSUM curve is that the energy-saving measures suddenly stopped being effective.
The more likely situation in practice would be a slow reduction in efficiency over a longer
period caused by, say the deterioration in equipment or poor maintenance. Over a short
period, the CUSUM curve for the successive implementation of two actions would be as
shown in Figure 6 with the savings due to each action as shown.
In the previous example, it was assumed that the energy-saving measures were first
implemented in week 9. In practice, the exact time would be logged in company records. If
such a company invests in energy-saving measures then they will expect a certain
percentage improvement. This expected performance is a target against which the actual
performance can be compared and both the anticipated and actual performance can easily
be plotted on a CUSUM curve on a regular weekly or monthly basis. In this way, CUSUM
is also a technique for monitoring. In the situation where no energy-saving measures have
been applied CUSUM will indicate any plant deterioration as a positive value above the
datum line. When remedial action is taken the CUSUM line will again return to the
horizontal, parallel to the datum line.
The technique has a wider application; for example, the heating requirements of a building
usually follow the pattern of degree days which apply to the locale. It can also be used
when energy consumption is related to more than one variable. In this case the regression
analysis for say, three variables may take the form:
E = A + (B x P1) + (C x P2) + (D x P3)
where the suffices relate to the production lines of products 1, 2, and 3, or more generally,
three independent variables.
The previous discussion has suggested that a linear relationship always exists between the
energy consumption and the independent variable(s). This is not always the case; certain
types of products are better described by non-linear energy consumption equations.
Providing there is enough reliable data, then all modern PC databases will have facilities
for undertaking the relevant regression analysis. The CUSUM technique simply requires
an appropriate formula for predicting the energy usage of a system.
Of course, the technique should be applied with care. Much depends on the accuracy of
the data. Unreliable data is revealed by the regression analysis not giving a near-horizontal
datum line as shown on Figure 6. In such cases the CUSUM analysis will be only a fair
general indicator of energy saving but a very good indicator of the need for more and
improved quality monitoring equipment. Equally, the situation where there is a product
mix on a single production line will require accurate data for each separate product within
the line if the CUSUM analysis is to be used.

Example of CUSUM

Foundry data
800000

700000

600000

kWh

500000

400000

300000

200000

100000

0
0

100

200

300

400

500

600

700

800

900

Tonnes

In this example the standing losses are small in comparison to the production related
energy use. The small negative intercept would be due to the scatter of the results.
Normally the intercept would be positive.

Degree days
Degree days provide a measure of the outside temperature and hence can be correlated to
the amount of energy required to heat or cool buildings.
Degree days can be simply defined as the product of the difference temperature and time,
that, the average outside temperature is below (heating) or above (cooling) a base
temperature.

Baseline graphs for buildings

energy

energy
1

degree days
energy

degree days
energy

degree days

degree days

energy
5

base 20.5 C
base 15.5 C
base 10.5 C

degree days

1. Typical relationship of the form Energy = m x DD + baseload (weather unrelated)

2. There is no weather-unrelated use ( e.g. an air heater in a large workshop).
3. The building is heated for specified months only (e.g. October - April). If the
temperature is below the base temperature outside of the "heating season" the
occupants go cold.
4. The heating system is not capable of meeting the heating load during the coldest
weather conditions. The building is not able to be heated to the required temperature
level.
5. It is necessary to choose the correct base temperature for the building.
For most domestic and commercial buildings the base temperature is 15.5C. Buildings
heated throughout the day such as hospitals have a higher base temperature 18.5C. Some
buildings may only require frost protection and it would be necessary to determine degree
days for a lower base temperatures e.g. 5C.

10