Professional Documents
Culture Documents
WO 2018/004464 Al: W!Po PCT
WO 2018/004464 Al: W!Po PCT
(51) International Patent Classification: DZ, EC, EE, EG, ES, FI, GB, GD, GE, GH, GM, GT, HN,
G05B 13/04 (2006.01) F24F 11/00 (2006.01) HR, HU, ID, IL, IN, IR, IS, JO, JP, KE, KG, KH, KN, KP,
KR, KW, KZ, LA, LC, LK, LR, LS, LU, LY, MA, MD, ME,
(21) International Application Number:
MG, MK, MN, MW, MX, MY, MZ, NA, NG, NI, NO, NZ,
PCT/SG20 17/050324
OM, PA, PE, PG, PH, PL, PT, QA, RO, RS, RU, RW, SA,
(22) International Filing Date: SC, SD, SE, SG, SK, SL, SM, ST, SV, SY, TH, TJ, TM, TN,
29 June 2017 (29.06.2017) TR, TT, TZ, UA, UG, US, UZ, VC, VN, ZA, ZM, ZW.
(25) Filing Language: English (84) Designated States (unless otherwise indicated, for every
kind of regional protection available): ARIPO (BW, GH,
(26) Publication Langi English
GM, KE, LR, LS, MW, MZ, NA, RW, SD, SL, ST, SZ, TZ,
(30) Priority Data: UG, ZM, ZW), Eurasian (AM, AZ, BY, KG, KZ, RU, TJ,
10201605346S 29 June 2016 (29.06.2016) SG TM), European (AL, AT, BE, BG, CH, CY, CZ, DE, DK,
EE, ES, FI, FR, GB, GR, HR, HU, IE, IS, IT, LT, LU, LV,
(71) Applicant: KIRKHAM GROUP PTE LTD [SG/SG]; MC, MK, MT, NL, NO, PL, PT, RO, RS, SE, SI, SK, SM,
1 Fusionopolis Walk, #02-1 1 North Tower, Singapore TR), OAPI (BF, BJ, CF, CG, CI, CM, GA, GN, GQ, GW,
138628 (SG). KM, ML, MR, NE, SN, TD, TG).
(72) Inventors: CHAI, Kok Soon; 385, Goodview Gardens,
Bukit Batok West Avenue 5, # 1 1-334, Singapore 650385 Declarations under Rule 4.17:
(SG). LAI, Choon Hoo; 25, Hazel Park Terrace, #18-02, — of inventorship (Rule 4.1 7(iv))
Singapore 678948 (SG).
Published:
(74) Agent: AMICA LAW LLC; 30 Raffles Place, #14-01 — with international search report (Art. 21(3))
Chevron House, Singapore 048622 (SG).
(81) Designated States (unless otherwise indicated, for every
kind of national protection available): AE, AG, AL, AM,
AO, AT, AU, AZ, BA, BB, BG, BH, BN, BR, BW, BY, BZ,
CA, CH, CL, CN, CO, CR, CU, CZ, DE, DJ, DK, DM, DO,
(54) Title: LARGE SCALE MACHINE LEARNING-BASED CHILLER PLANTS MODELING, OPTIMIZATION AND DIAG-
NOSIS
(57) Abstract: The invention relates to a data driven, or a hybrid rule-based
and data driven Energy/Building Management System, such as for chiller plants,
which has ability to learn from the data and evaluate performance. According
to the invention, a computer-implemented method trains prediction models for
each equipment model and chiller plant model using baseline data, predicts a
parameter for each equipment model and chiller plant model using baseline da
ta, computes differential parameter of each equipment based on the predicted
and actual parameters of each equipment, computes differential parameter of the
chiller plant based on the predicted and actual parameters of the chiller plant,
compute a differential parameter resulting from chiller plant optimization, by
subtracting the differential parameters of the various equipment from the differ
ential parameter of the chiller plant,ascertaining a presence of abnormality in the
differential parameter resulting from chiller plant optimization and generating a
notification if the differential parameter resulting from chiller plant optimization
is ascertained abnormal.
o
© 417
00
o
Large Scale Machine Learning-based Chiller Plants Modeling, Optimization and
Diagnosis
Field of Invention
Background
Chiller plant optimization is one of the most crucial tasks to smart building systems,
as the energy consumption of a chiller plant comprises over 40% of the total energy
consumption of a modern building. Poor efficiency is commonly observed in existing chiller
plant systems, due to the excessive overhead and technical challenges faced in manual
tuning. In practice, a large number of chiller plants and buildings are optimized during the
first few months in operation, when experienced engineers spend huge efforts on fine-tuning
the chiller plants to achieve near-optimal performance. However, the efficiency of these
chiller plants deteriorates quickly when the engineers with expertise leave the projects, such
that the configuration of the chiller plant does not adapt well to the varying environmental
and equipment conditions. A fully optimized chiller plant may run at excellent efficiency
during office hours, but performs poorly at nights/weekends/public holidays. Moreover, even
veterans in the industry may not always make correct decisions on chiller plant optimization.
The extremely high complexity of chiller plants often leads to ineffectiveness of conventional
tuning tricks used by the engineers. The best engineers may only tune the chiller plant
system in a trial-and-error fashion, trying to understand the chiller plants with their
experience and sometimes simple heuristics.
l
plant. The model is also a best fit for the data for a relatively short period of time, but it loses
predicting accuracy over new data set after a period of chiller plant operation.
Reference is made to [1] and [2]. In [1], Nguyen developed Table 1 to describe the
existing HVAC optimization process that is divided into Preprocessing, Running Optimization
and Post-processing. In [2], neural network was applied to air conditioning and chiller plants
systems.
[1], [2] and other existing systems focus on model development and system
optimization for one specific chiller plant and for relatively short period of time. For any of the
existing systems to be suitable as an industrial solution, it will face major challenges to meet
"scalability, adaptability and continuous learning" requirements.
Summary
One aspect of the invention covers the application of a method and system to use
data driven model and sub-models, e.g. neural networks with inputs layers, multiple hidden
layers, and output layers, to represent the actual equipment and predict equipment and
chiller plant power, performance and efficiency. The system learns to represent the chiller
plant and equipment high level characteristics from relative accuracy of the predicted values.
It further learns to represent detailed equipment characteristics such as but not limited to
equipment efficiency, performance etc. to provide data driven, detailed and actionable
diagnostics information for further analysis, diagnostics or energy efficiency optimization.
Another aspect of the invention covers the application of the data driven and deep
learning with model and sub-models to predict and evaluate efficiency and performance of
the system and various equipment. The model and sub-models are trained to predict system
and equipment power and performance using trained data, X train , and evaluated with a set of
accuracy matrix, Macc . The model and sub-models are evaluated with a set of cross
validation data, X cr0S s, for accuracy evaluation. By evaluating and meeting certain accuracy
requirements in Macc , the model and sub-models are used as the baseline for predicting
power, performance and efficiency from the M&V data, X test, of the equipment and of chiller
plant in the future. The model and sub-models predict the performance values of the system
and equipment to be used for comparison for the actual performance values. The system
refers to the accuracy matrix and variance in performance of system and subsystems
between the training Xtrain , cross validation X cross and testing xte st data sets to conclude
performance evaluation of the system and subsystem over time.
Another aspect of the invention covers a method and system that refers to the
performance matrix of equipment for large scale chiller plant comparisons and cross learning.
Another aspect of the invention covers a data driven "universal" model that
generalizes the system and equipment performance, power and efficiency using the
measurement and verification sensor data of the chiller plants that the model is trained on,
as well as the cross validation data and testing data that the model have not been trained on.
This includes a method and system to achieve the best trade-off for optimizing the
generalization of any chiller plants. It includes a method to model the chiller plant life cycle
for continuous and automated learning, modelling and optimization.
training a plurality of prediction models using first baseline data, the prediction
models being for chiller plant and a plurality of equipment comprising cooling tower
(CT), condenser water pump (CWP), chiller, chilled water pump (CHWP);
computing, using the prediction models, a plurality of predicted parameters of
the plurality of equipment and the chiller plant using test data;
computing a plurality of differential parameters of the plurality of equipment
based on the predicted parameters of the plurality of equipment and a plurality of
actual parameters of the plurality of equipment;
computing a differential parameter of the chiller plant based on the predicted
parameter of the chiller plant and an actual parameter of the chiller plant;
computing a differential parameter resulting from chiller plant optimization, by
subtracting the differential parameters of the plurality of equipment from the
differential parameter of the chiller plant;
training one of the prediction models, which corresponds to the any one of the
plurality of equipment, using second baseline data.
Figure 2A shows a machine learning life cycle that addresses the challenges of
scalability, adaptability and learning;
Figure 2B shows a basic concept of generalization of model for chiller plants, and
four phases required to build a model that achieves universal fitting of chiller plant data;
Figure 3B shows prediction models for chiller plant equipment e.g. cooling tower,
condenser water pump, chiller model, chilled water pump, using an equipment performance
decomposition-based approach with increasing levels of abstraction;
Figure 3C shows one embodiment of prediction models for chiller plant equipment,
e.g. cooling tower, condenser water pump, chiller model, chilled water pump;
Figure 3F shows one example of a prediction model for condenser water pump;
Figure 3G shows one example of a prediction model for chilled water pump;
Figure 5 shows that the LEO system provides visualization data with targeted areas
for human intervention and automatic optimization as compared to existing HVAC Big Data
Tools;
Figure 6 shows that LEO system performs machine and statistics machine learning
to deliver two types of results;
Figure 8 shows introduction of high variances to training data for cooling tower by the
invention;
Figure 9 shows a comparison of predicted total chiller plant power, actual total chiller
plant power and a deviation therebetween;
In the following description, numerous specific details are set forth in order to provide
As used herein, the terms "first," "second," and "third," etc. are used merely as labels,
and are not intended to impose numerical requirements on their objects.
As used herein, the term "power" includes references to "power consumption" and
may be interchangeably used; the terms "model" and "sub-model" include references to
"prediction models" and "machine learning models", and may be interchangeably used; the
terms "cooling load" includes references to "cooling tonnage" and may be interchangeably
used.
Figure 1 shows the system architecture of LEO system that receives real time M&V
(measurement and verification) data from chiller plant, or historical data from different
sources to trigger continuous machine learning life cycles. A machine learning life cycle
includes preprocessing, optimization and post-processing.
Figure 2A shows the novel machine learning life cycle that addresses the challenges
of scalability, adaptability and learning. The machine learning life cycle is divided into two
major phases, namely i) model training phase and ii) large scale machine learning-based
prediction and optimization phase.
Model training phase (in blocks 201 a and 201 b of Figure 2A)
The machine learning phase involves the development of a "universal model" that
generalizes by fitting accurately to the measurement and verification data of a smaller
sample of chiller plants. In particularly, the model comprises of sub-models that represent
performance, power and efficiency of equipment such as chillers, chilled water and
condenser water pumps, cooling towers. The equipment list can be extended to airside
equipment such as air handling units and fan coil units etc. The main objective in the training
phase is to achieve the best fit for a relative small sample of chiller plants with the M&V data
t-Phase,
while satisfying the best fit for large scale of chiller plants' M&V data that the model
is not trained to fit in the training phase. In summary, the "universal model", is trained using
the Xt-phase data, but it is expected to predict performance, power, performance and efficiency
of future M&V data X m i-Phase without reprogramming. Deep neural network and multi-level
regression model would likely be the best way to build the universal model. In Ref [3], Wei
describes a data driven method to model a chiller plant. Wei applies BFGS (broydene-
fletcheregoldfarbeshanno) method to a one-output-unit MLP (multi-layer perceptron) to train
a network to represent the prediction model. Monfet and Lee also applied neural network to
represent prediction models that predict single output. These models are black box
approaches that predict an output, e.g. chiller plant power or efficiency with multiple inputs.
The invention applies a novel deep learning approach that decomposes a chiller plant model
to multiple sub-models e.g. equipment models (see Figure 3B), to form a final prediction
model.
Other preprocessing tasks will also be performed in the model training phase
[4],[5],[6].
Large scale machine learning phase (in blocks 202 to 208 of Figure 2A .
The LEO system with "universal model" is ready for large scale machine learning in
this phase. All the tasks performed in the large scale machine learning phase are automated,
and LEO system will only prompt for user intervention if it detects abnormal sensor data that
is beyond uncertainty levels. The large scale learning phase is divided into 5 tasks:
a) Automatic model training and fine tuning for a specific chiller plant (in block
202) . The LEO receives sensor data from different M&V data sources, X m
Phase, where ml-phase refers to machine learning phase, and automatically
trains features and its weights to accurately fit the data, X m i-Phase- E system
shall initiate a set of performance and accuracy diagnostics for sensor data
accuracy and uncertainty. Figure 2B shows the novel framework that LEO
system applies for the development of the prediction model for the chiller
plant. It further applies the novel equipment decomposition-based model
development approach to improve the speed and accuracy of model training.
b) Optimization (in block 203) . It shall trigger an optimization program to
determine the optimum values of certain set points, SP m i- P h ase that minimize
total chiller plant power consumption. Block 203 may be performed using
existing methods.
c) Efficiency evaluation and sensitivity analysis (in block 204) . LEO system will
evaluate the actual efficiency of the chiller plant and uncertainties of the
sensor data to determine the fitness of the universal model. It decides to
trigger a bottom up, localized model retraining that will evaluate performance
and efficiency of individual equipment. Equations 5 and 6 are applied as a
novel differential equipment performance approach to evaluate the equipment,
and chiller plant efficiency.
d) Story Development from the raw data (in block 205) . LEO system will apply
the differential equipment performance approach to develop a story with time
line and descriptions on the changes of equipment and chiller plant
performance and energy efficiency. LEO recursively splits the raw data into
different time periods based on the differential equipment performance
approach.
e) Retrain the model (in block 206) . The invention applies the concept of
differential equipment performance to evaluate i) the effectiveness of energy
efficiency optimization algorithm, ii) the deviation of equipment performance
and efficiency and decide how to retrain the universal model to improve
fitness to the M&V data, and manage uncertainty.
f) Localized machine learning in 207 . The task is started once the model's
fitness to the M&V data is in question. The localized machine learning is a
localized, sub-model training approach that trains for fitness of the data to
individual equipment
g) Stop training 208 . The task is started once the uncertainty of the data
accuracy exceeds certain level, and the sensor data is not fit for input to the
model for optimization. A user is informed to fix the sensor data before
machine learning can resume optimization .
Figure 2B shows the basic concept of generalization of model for chiller plants. The
"universal model" is a multi-model based on a hybrid of deep neural network supported by
continuous machine learning life cycles. The novelty is in model development process as
well as the continuous machine learning life cycles. Figure 2B shows that 4 phases are
required to build a model that achieves universal fitting of chiller plant data.
Chiller plant M&V data can be modeled as inputs x . . x n to be associated with one or
more outputs y (see Figure 3A). The model would learn a set of weights w . -w and
compute their outputs f(x,w) .
There are multiple requirements to be met to achieve a universal model. The first
requirement is an optimization problem to minimize total chiller plant power vs total chiller
plant cooling tonnage, i.e.
Min { Rtotal } .
decomposition-based approach to represent a prediction model that will meet the second
requirement. The third requirement is a Phase 3 requirement that would be met by modeling
special features that detect changes in the life cycle of a chiller plant.
Figure 3A and 3B show a model and sub-models for learning to represent chiller
plants with a similar architecture with different level of abstractions. The total equipment
power of a chiller plant is P = P + P + P + P + P . Formally, the sub-models
total chwp chiller cwp ct air
are trained to represent the f(x) : R → R , where D is the size of input vector X and L is the
size of the output vector f(x). In particularly, the f(x) Chw P for block 309 in Figure 3B represents
the power and efficiency for chilled water pumps with data sets of flow rates and other
features as inputs for training. The f(x) cwp block 307 in figure 3B also represents the power
and efficiency for condenser pumps using features such as flow rates as inputs.
The f(x) ct and f(x) Chiiier that represent the power and efficiency of cooling towers and
chillers are significantly more difficult to be modelled. The motivation of the algorithm is to
learn to model the power and efficiency of the cooling towers and chillers to trade off better
ratios of total power over total cooling tonnage.
For chillers, chiller power consumption f(x) Chiiier for block 308 in figure 3B can be
modelled by independent parameters stated in the chiller manufacturing datasheet such as
chilled water supply set point, condenser supply temperature, usgpm/rt (flow rate) for chilled
water and condenser water, and the cooling tonnage supplied by the chiller. These
parameters are used to train with the chiller's power consumption to obtain the f(x) Chiiier-
For cooling tower, the problem is constructed such that given some noisy
observations of a dependent variable at certain values of the independent variables {wet
bulb, dry bulb, rh, cooling load and may other variables}, what is the best estimate of the
dependent variable at a new value, f(x) ct for block 306 in figure 3B.
P + P + P + P + P (equation 1)
chwp chiller cwp ct air
E = E + E + E + E + E (equation 2)
total chwp chiller cwp ct air
P o a
i refers to total equipment power, Pchwp refers to chilled water pump power, P chi ne refers
to chiller power, Pcwp refers to condenser water pump power, Pc, refers to cooling tower
power, Pai refers to air side equipment power.
E total refers to equipment efficiency, E chwp refers to chilled water pump efficiency, Echi iier refers
to chiller efficiency, Ecwp refers to condenser water pump efficiency, Ec, refers to cooling
tower efficiency, Eai refers to air-side equipment efficiency.
A set of prediction models are provided for a chiller plant and a plurality of equipment
comprising cooling tower, condenser water pump, chiller and chilled water pump. These
models are configured to train or machine learn from baseline data and thereafter predict
parameters, e.g. power, flow and/or temperature, for their respective chiller plant or
equipment during test period.
Figure 3A shows a prediction model for chiller plant, e.g. chiller plant model 320,
which is configured to predict chiller plant power based on weather and cooling load (RT).
Weather and cooling load are independent variables.
In Figure 3C, a first prediction model 306 is configured to predict a condenser water
temperature into chiller (cwshdr); a second prediction model 307 configured to predict a
condenser water flow in/out chiller (cwfhdr); a third prediction model 309 is configured to
predict a chilled water flow in/out chiller (chfhdr); and a fourth prediction model 308 is
configured to predict a chiller power (chkw) based on the condenser water temperature into
chiller (cwshdr), the condenser water flow in/out chiller (cwfhdr), the chilled water flow in/out
chiller (chfhdr), a cooling load and a chiller set-point (chsp).
In Figure 3D, the first prediction model 306 includes a first and a second sub-model.
The first sub-model 306a, e.g. cooling tower model (CT), which is for cooling tower
equipment is configured to predict cooling tower power ("ctkw"), e.g. in kilowatts, based on
VSD (variable speed drive) speed of cooling tower fan ("ct_speed"). The second sub-model
306b, e.g. condenser water temperature model (CWTM), which is for condenser water pump
equipment, is configured to predict condenser water temperature into chillers ("cwshdr")
based on weather and VSD (variable speed drive) speed of cooling tower fan ("ct_speed").
In Figure 3D, the second prediction model 307 includes a third and a fourth s ub
model. The third sub-model 307a, e.g. condenser water flow model (CWFM), which is for
condenser water pump equipment, is configured to predict condenser water flow in/out of
chillers ("cwfhdr") based on VSD speed of condenser water pump ("cwp speed"). The fourth
sub-model 307b, e.g. condenser water pump model (CWP), which is for condenser water
pump equipment, is configured to predict condenser water pump power ("cwpkw") based on
VSD speed of condenser water pump ("cwp speed").
In Figure 3D, the third prediction model 309 includes a fifth and a sixth sub-model.
The fifth sub-model 309a, e.g. chilled water pump model (CHWP), which is for chilled water
pump equipment, is configured to predict chilled water pump power ("chwpkw") based on
VSD speed of chilled water pump ("chwp speed"). The sixth sub-model 309b, e.g. chilled
water flow sub- model (CHFM), which is for chilled water pump equipment, is configured to
predict chilled water flow in/out of chillers ("chfhdr") based on VSD speed of chilled water
pump ("chwp speed").
In Figure 3D, the fourth prediction model 308, e.g. chiller model 308, which is for
chiller equipment, is configured to predict chiller power ("chkw") based on condenser water
flow in/out of chillers ("cwfhdr"), condenser water temperature into chillers ("cwshdr"), chilled
water flow in/out of chillers ("chfhdr"), cooling load and chilled water set point ("chsp").
Cooling load and chilled water set point ("chsp") are independent variables.
The set of prediction models may further comprise a fifth prediction model 310, e.g
chiller plant equipment model (see Figure 3B), which is configured to predict total equipment
power based on cooling tower power ("ctkw"), condenser water pump power ("cwpkw"),
chiller power ("chkw") and chilled water pump power ("chwpkw") from prediction models 306,
307, 308, 309.
Reference is made to Figure 3E which shows another example of the first prediction
model of Figure 3C. The first prediction model 306 of Figure 3E is configured to predict
condenser water temperature into chiller (cwshdr) based on weather data, cooling load,
cooling tower power (ctkw).
Reference is made to Figure 3G which shows another example of the third prediction
model of Figure 3C.The third prediction model 309 of Figure 3E is configured to predict
chilled water flow in/out chiller (chfhdr) based on chilled water pump power (chwpkw).
In some embodiments, any or all of the first, second, third prediction models of Figure
3D may be replaced by the respective model of Figures 3E, 3F and 3G, and the predicted
parameters would be modified according to the replacement models as described in relation
to Figures 3E, 3F and 3G.
In some other embodiments, that all of the first, second, third prediction models of
Figures 3E, 3F and 3G may be combined with the fourth prediction model 308 to arrive at
Figure 3H which shows another embodiment of the prediction model of Figure 3C. The
predicted parameters would be modified according to the replacement models as described
in relation to Figures 3E, 3F and 3G.
Phase 3 - Time based model training using baseline, test and cross-validation data
Performance evaluation process for a chiller plant to meet phase 3's requirement in
Figure 2B is explained below.
Performance evaluation is based on the concept of comparing the actual present
power (Pow present ) versus the actual historical power (Pow historica i) at similar conditions. The
conditions are the reverse projection of the actual power consumption to the prediction
features in blocks 306-309 in figure 3B. This can be achieved using a couple of methods, e.g.
deep learning models that estimate the relationship between dependent variables (blocks
306-309), e.g. powers, and independent variables, or features (blocks 302-309), e.g. flow
rates, temperatures and powers. The chiller plant level performance comparing a period of
present time t(pres,t) = {t(pres,1) .... T(pres,n)} to a period of historical time t(his,t) =
{t(his,1)... t(his,m)}. Equipment performance decomposition-based approach in Figure 3 B
builds a prediction model that estimates the equipment and plant level dependent variables,
e.g. power consumption, using some given independent variables.
a) Equipment evaluation.
Differential power is computed by comparing Pow p eSent d Pow h iStoricai using the
following equation:
In block 401 , an operator decides what the objective of the prediction process is.
There may be two different objectives in the prediction process - i) objective 1 - Continuous
chiller plant representation learning and automatic optimization, ii) objective 2 - Story telling
from the M&V data for diagnostics.
In block 403, for objective 1, LEO divides raw or existing M&V data from sensors into
different time periods, and classifies them as baseline data, cross-validation data and test
data. For objective 2 , LEO classifies the M&V data into a baseline, a cross validation and a
number of (or at least one) test periods. Accordingly, data from baseline time period may be
referred to as "baseline data"; data from cross-validation time period may be referred to as
"cross-validation data"; data from test time period may be referred to as "test data". It is to be
appreciated that the cross-validation time period may be a subset of the baseline time period.
It is to be appreciated that the baseline time period and the test time period may or may not
be mutually exclusive.
In block 405, prediction models for equipment 306a, 306b, 307a, 307b, 308, 309a,
309b are trained to predict equipment power using baseline data. Training may be
performed by neural network model, Gaussian process, or other suitable methods, to
produce trained prediction models, e.g. cooling tower model 306a to predict cooling tower
power, condenser water pump model 306b, 307a, 307b to predict condenser water pump
power, chilled water pump model 309a, 309b to predict chilled water pump power, chiller
model 308 to predict chiller power (see blocks 407a, 407b, 407c, 407d). Similarly, prediction
model 320 for chiller plant is trained using baseline data to produce a trained chiller plant
model 320 to predict chiller plant power.
In blocks 409a, 409b and 409c, the trained prediction models, after training in block
405, are used to predict power for each equipment and for the selected time periods of block
403. As illustrated in block 409a, using baseline data, a predicted power is computed for
each equipment. As illustrated in block 409b, using cross-validation data, a predicted power
is computed for each equipment. As illustrated in block 409c, using test data, a predicted
power is computed for each equipment. Similarly, predicted power is computed for chiller
plant separately, using baseline, cross-validation and test data.
The predicted power computed using baseline data (see block 409a) may be
compared with the predicted power computed using cross-validation data (see block 409b)
to compute a deviation. For example, for cooling tower, predicted power computed using
baseline data is compared with predicted power computed using cross-validation data to
compute a deviation therebetween. Similar comparison and/or computation of deviation is
also performed for each of the remaining equipment i.e. chilled water pump, condenser
water pump, chiller, as well as for chiller plant.
In block 4 1 3 , model accuracy or fitness e.g. in terms of resolving bias and overfitting,
is analysed for each equipment based on the above-computed deviation between predicted
power using baseline data and predicted power using cross-validation data. For example,
for cooling tower, if the computed deviation falls within a predetermined threshold or limit,
accuracy of the cooling tower model is validated. However, if the computed deviation
exceeds the predetermined threshold or limit, the cooling tower model is not accurate or is
unfit and may require re-training of the cooling tower model. If re-training of the cooling
tower model is required, the process proceeds or returns to block 403 where a different
baseline data is to be selected and training of the cooling tower model takes place based on
the selected different baseline data.
22 nd March is higher while flow rate on 28th March is lower. An increased deviation of
around 24%, from baseline value, on 28th March infers a presence of abnormality in the
condenser water pump. By using scatter diagram, or scatter plot, the plot uses condenser
flow rates to determine power consumption of the condenser pumps.
To achieve objective 2 , the LEO system may perform recursive searches to identify
the change in equipment performance (block 4 1 5) to infer the root causes for the change of
chiller plant efficiency. The recursive searches may be performed by splitting the raw data
into multiple time periods, repeating the computations of block 4 1 5 , and comparing
equipment performance and/or efficiency over these multiple time periods.
Block 417 evaluates the effectiveness of chiller plant optimization. To evaluate the
effectiveness, e.g. improvement or decline, of chiller plant optimization, a differential power
or change in power consumption contributed by chiller plant optimization is computed. To
this end, the differential powers of total equipment and of chiller plant are computed.
To compute differential power of the chiller plant, a predicted power computed from
test data, using the chiller plant model 320, is compared with actual power of chiller plant
during test period to compute a difference therebetween.
From block 4 1 7 , the total change of equipment performance and/or efficiency may be
used to isolate the results of optimum control strategies, and the change of sensor
accuracies.
For example, if the computed differential power resulting from chiller plant optimization
implies an improvement in chiller plant operation, e.g. the computed differential power
resulting from chiller plant optimization is a negative value or becomes an increasingly
negative value over a period of time, this shows that power consumption has decreased,
chiller plant optimization has been effective and therefore human intervention may not be
required. However, if the computed differential power resulting from chiller plant optimization
implies a decline in chiller plant operation, e.g. breaches a predetermined threshold, a
presence of abnormality is ascertained and a notification is generated and provided to an
operator to request human intervention, e.g. manual check on physical equipment and/or
optimization strategies. This predetermined threshold may be defined as: the computed
differential power resulting from chiller plant optimization is a positive value greater than a
predetermined value, or becomes an increasingly positive value over a period of time, etc. If
abnormality in the computed differential power resulting from chiller plant optimization is
ascertained present, the process may proceed to block 405 to re-train the equipment models.
When the computed differential power resulting from chiller plant optimization (block
4 1 7) is considered together with the differential powers of each or total equipment (block
4 1 5), an improvement in their differential powers implies that chiller plant optimization has
been effective and therefore human intervention may not be required; conversely, a decline
in their differential powers implies presence of abnormality in chiller plant optimization and/or
The foregoing paragraphs describing the process of Figure 4 are based on the
prediction models of Figure 3D, where the power parameters are predicted and compared. It
is to be appreciated that if the process illustrated in Figure 4 is based on the prediction
models of Figure 3H, or based on the prediction models of Figure 3D being replaced, in part,
by any of the prediction models of Figure 3E, 3F and 3F, the foregoing computation,
comparison and/or analysis steps would be suitably modified. Particularly, as the first
prediction model 306 of Figure 3E predicts a temperature parameter, the second prediction
model 307 of Figure 3F predicts a flow parameter, the third prediction model 309 of Figure
3G predicts a flow parameter while the fourth prediction model 308 for chiller predicts a
power parameter, the steps described in relation to blocks 4 1 1, 4 1 3 , 4 1 5 and/or 4 1 7 will be
performed based on differential flow parameters, differential temperature and/or differential
power. Alternatively, flow and/or temperature parameters may be converted to power
parameter as known to person skilled in the art during performance of the steps of blocks
4 1 1, 413, 4 15 and/or 4 17 .
The lack of generality in the data is an important problem of existing systems, which
does not attract attention in almost all existing studies on chiller plants. Simple data
modelling over the existing chiller plant data may result in useless model with high
generalization error. In an extreme case, a chiller plant always runs at a fixed configuration,
e.g., fixed VSD speed for pumps and fans. By training data from this chiller plant, the result
ing data model is only applicable to the current configuration, and does not generate
meaningful prediction for any other configuration. Figure 8 plots the data distribution over
cooling tower speed (CT Speed) and cooling tower power, collected in fully controlled chiller
plant with a fixed VSD configuration setting (denoted as original data 801) and random VSD
configuration (denoted as rich data 802) respectively. The cooling tower fan is mainly
operated at the speed between 20% to 40% of the maximum speed. The results show that
data model using fixed VSD configuration does not have much generalization capability
when other configurations are used by the chiller plant whereas the invention enables
training based on a wider data variation.
Existing tools such as Lucid and Green Koncepts apply big data technologies to
generate nice visualization charts and data. However, better classification of energy data is
not sufficient and it still requires intensive and real time analysis to translate data into actions.
Energy management systems must advance from providing charts and data to automatic
optimization and targeted diagnostics information. The users of energy management
systems are technicians and facility managers who do not have time and expertise to
develop advanced machine learning techniques to translate data and charts to actions that
improve energy efficiency.
Figure 5 shows that the LEO system provides visualization data with targeted areas
for human intervention and automatic optimization as compared to existing HVAC Big Data
Tools.
Figure 6 shows that LEO system performs machine and statistics machine learning
to deliver 2 types of results: i) Automatic optimization that results in improved energy
efficiency, ii) Provide some very specific actionable information for users to take specific
actions. For example, LEO system suggests that the energy efficiency of the chiller 3 is
gradually going down by x.xx% since it was serviced 3 months ago. LEO system predicts
that restoring the energy efficiency of the chiller to its original conditions would result in a
saving of $xxx per months.
The in-premise control system 1201 includes sensors and/or actuators 1203 which
are communicably coupled to chiller plant equipment for sensing data parameters, e.g.
measure speed, flow, temperature and/or power, of equipment and/or for controlling
equipment.
The in-premise control 1201 further includes the LEO system 1204 (see also Figure 1)
which is configured to perform real-time learning, control, optimization and diagnostics. The
LEO system 1204 includes at least one memory storage, and at least one computing unit
communicably coupled thereto. The memory storage is configured to store data (e.g.
baseline data, cross-validation data, test data), prediction models, and computer-executable
instructions that, when executed by the at least one computing unit, cause performance of
operations as described in the present disclosure and in relation to blocks 401 to 417. The
computing unit of the LEO system may be communicably coupled to the sensors and/or
actuators 1203, directly or indirectly, to receive measured data transmitted by the sensors
and/or actuators 1203.
The LEO system 1204 is communicably coupled to the cloud server 1202 for data
transmission therebetween, and to at least an input/output device, e.g. display unit and data
entry unit, to provide a user interface. The user interface or display unit is configured to
display notifications and/or diagnostics from the LEO system 1205, and allow viewing of
reports generated by the LEO system 1205. The user interface 1206 or input unit is further
configured to allow an operator, e.g. building and/or database manager, provide instructions
to control equipment and/or chiller plant.
Other embodiments will be apparent to those skilled in the art from consideration of
the specification and practice of the invention. Furthermore, certain terminology has been
used for the purposes of descriptive clarity, and not to limit the disclosed embodiments of the
invention. The embodiments and features described above should be considered exemplary.
List of References
[1] Nguyen, A . T.; Reiter, S.; Rigo, P. A review on simulation-based optimization methods
applied to building performance analysis. Applied Energy 1 13 (2014) 1043-1058.
[2] Applications of artificial neural networks for refrigeration, air-conditioning and heat pump
systems— A review, RENEWABLE AND SUSTAINABLE ENERGY REVIEWS FEBRUARY
2012.
[3] Xiupeng Wei, Guanglin Xu, Andrew Kusiak, "Modeling and optimization of a chiller plant",
Energy, 2014, Elsevier.
[4] Vakiloroaya, V., Q . P. Ha, and B . Samali. "Energy-efficient HVAC systems: Simulation-
empirical modelling and gradient optimization." Automation in Construction 3 1 (2013): 176-
185.
[5] D. Monfet, R . Zmeureanu, Ongoing commissioning of water-cooled electric chillers using
benchmarking models, Applied Energy 92 (201 2) 99-108.
[6] Lee, T.-S.; Liao, K.-Y.; Lu, W.-C. Evaluation of the suitability of empirically-based
models for predicting energy performance of centrifugal water chillers with variable chilled
water flow. Appl. Energy 201 2 , 93, 583-595.
[7] Hartman, Thomas. "All-variable speed centrifugal chiller plants." ASHRAE journal 43.9
(2001): 43.
[8] Yu, F. W., and K . T . Chan. "Optimization of water-cooled chiller system with load-based
speed control." Applied Energy 85.10 (2008): 931 -950.
[9] Kohavi, Ron. "A study of cross-validation and bootstrap for accuracy estimation and
model selection." Ijcai. Vol. 14. No. 2 . 1995.
Claims
if the presence of abnormality in the differential parameter resulting from chiller plant
optimization is ascertained, generating a first notification which identifies a request for
human intervention.
if the presence of abnormality in any one of the differential parameters of the plurality
of equipment is ascertained, performing at least one of the following steps:
training one of the prediction models, which corresponds to the any one of the
plurality of equipment, using second baseline data.
3 . The method of any one of claims 1 to 2 , further comprising:
computing a first plurality of deviations between the predicted parameters of the
plurality of equipment using the baseline data and the predicted parameters of the plurality of
equipment using cross-validated data;
computing a second deviation between the predicted parameter of the chiller plant
using the baseline data and the predicted parameter of the chiller plant using cross-
validation data; and
ascertaining accuracy of the prediction models based on the first deviations and the
second deviation.
5 . The method of any one of claims 1 to 4 , wherein the prediction models include:
a first prediction model configured to predict a condenser water temperature into
chiller (cwshdr);
a second prediction model configured to predict a condenser water flow in/out chiller
(cwfhdr);
a third prediction model configured to predict a chilled water flow in/out chiller
(chfhdr); and
a fourth prediction model configured to predict chiller power (chkw) based on the
condenser water temperature into chiller (cwshdr), the condenser water flow in/out chiller
(cwfhdr), the chilled water flow in/out chiller (chfhdr), a cooling load and a chiller set-point
(chsp).
9 . The method of any one of claims 1 to 5 and 8 , wherein the parameters include power,
flow and temperature.
if the presence of abnormality in any one of the differential parameters of the plurality
of equipment is ascertained, performing at least one of the following steps:
training one of the prediction models, which corresponds to the any one of the
plurality of equipment, using second baseline data.
12. The system of any one of claims 10 to 11, wherein the operations further comprising:
computing a first plurality of deviations between the predicted parameters of the
plurality of equipment using the baseline data and the predicted parameters of the plurality of
equipment using cross-validated data;
computing a second deviation between the predicted parameter of the chiller plant
using the baseline data and the predicted parameter of the chiller plant using cross-
validation data; and
ascertaining accuracy of the prediction models based on the first deviations and the
second deviation.
14. The system of any one of claims 10 to 13, wherein the prediction models include:
a first prediction model configured to predict a condenser water temperature into
chiller (cwshdr);
a second prediction model configured to predict a condenser water flow in/out chiller
(cwfhdr);
a third prediction model configured to predict a chilled water flow in/out chiller
(chfhdr); and
a fourth prediction model configured to predict chiller power (chkw) based on the
condenser water temperature into chiller (cwshdr), the condenser water flow in/out chiller
(cwfhdr), the chilled water flow in/out chiller (chfhdr), a cooling load and a chiller set-point
(chsp).
16. The system of any of claims 10 to 15, wherein the parameters include power.
18. The system of any one of claims 10 to 14 and 17, wherein the parameters include power,
flow and temperature.
PCT/SG201 7/050324
B. FIELDS SEARCHED
Minimum documentation searched (classification system followed by classification symbols)
G05B, F24F
Documentation searched other than minimum documentation to the extent that such documents are included in the fields searched
Electronic data base consulted during the international search (name of data base and, where practicable, search terms used)
FAMPAT: model, chiller, plant, equipment, predicted, actual, differential, subtract, abnormality, optimization, condenser,
¾ , fti†, ¾ , ¾ , , and related terms.
| Further documents are listed in the continuation of Box C . | See patent family annex.
Date of the actual completion of the international search Date of mailing of the international search report
0 1/08/201 7 . ,
27/07/201 7 (day/month/year) (day/month/year)
Singapore 189554
Email: pct@ipos.gov.sg IPOS Customer Service Tel. No. : (+65) 6339 861 6
Note: This Annex lists known patent family members relating to the patent documents cited in this International Search
Report. This Authority is in no way liable for these particulars which are merely given for the purpose of information.
JP 201221 0 15 1 A 25/10/201 2
SG 179328 A 1 27/04/201 2