AI and Deep Learning for Chiller Plant Optimization

1.    Introduction

Energy efficient design is one of several reactions of the engineering community towards a more sustainable and human oriented world. But this doesn’t mean only to design systems with high performing equipment (EER, COP etc.). There is also need to create tools that will easily deployed into existing systems and make them more energy efficient, self-learning and autonomously optimized without human interference.

Artificial Intelligence and Machine Learning is that field of computer science that allows to build intelligent entities which will interact with their environment, learn from experience and act on the most optimum way towards a certain performance task. 

Scope of current blog is to present a holistic approach of how to implement new developed Machine Learning (ML) technics towards an optimum and efficient operation of existing chiller plants.

The objective is to model the operation of a chiller plant in a way so that input data from the system itself under normal and continuous operation will be processed, vectorized and be fed into a deep learning neural network which will be trained to predict the system performance on a real time base.

By this way the system can be self-learning and have an integrated ML platform which will perform its own different system configuration-trials before any action taken from the system controller (BMS, SCADA, PLC). Based on very fast set-up and plant configuration checks and evaluating the predicted optimum performance, the ML agent will pick and apply the most highly performing settings to the system operation.

2. Structure of the Learning Agent

According to S.Russell and P.Norvig [1], the general definition of a learning agent is:

 An Agent is anything that can be viewed as perceiving its environment through sensors and acting upon the environment through actuators 

In our case the learning agent within a chiller plant can be a software agent which receives network packets as sensory inputs and acts on the environment by sending network packets to actuators.

2.1 Agent Perception and Action

Obviously our agent continuously tries to have a perception of its environment. Elements like the environmental conditions (ambient temperature, humidity), the equipment operation status (on/off, power demand, return/supply water temperatures etc.) along with the terminal load cooling demand constitute the environment of the agent. The learning agent is an intelligent entity which continuously interact with all elements of the aforementioned environment by two ways:

• Using sensors (temperature, humidity, electrical energy measurement, thermal energy measurement, water flow, etc.) to perceive the existing status of all elements of its considered environment.
• Proceed into actions through actuators (equipment on/off signals, pumps and funs speed regulation, motorized valves actuation, etc.) and make the system perform on an optimum way based on current environment status.

The same process happens iteratively and the holistic agent’s behavior is described by the agent function that maps any given percept sequence to an action.

A visual representation of the agent, the environment limits and the flow of information through sensors and actuators is as below.

Fig.01 - The Learning Agent

2.2 Definition of Agent behavior

Any approach to model the behavior of a chiller plant is a very complex procedure which has to consider the operation parameters of all individual equipment. The description of values like temperatures, flows, power and pressures are all expressions of second and third order mathematical equations.

 Also the interactions between several plant equipment and the influence of system or ambient conditions make the model even more complex and difficult to predict. It becomes a non-linear mathematical description which demands large number of assumptions to be solved. Any potential results are not accurate since critical relations of the system parameters are hidden behind non-linearities.

Instead of trying to model every specific chiller behavior through extended mathematical expressions the objective becomes on how to build AI learning agents and then to teach them.

The essential elements of a learning agent and the respective information flow can be visualized as below.

Fig.02 - Structure of the Learning Agent

According to S.Russell and P.Norvig [1] a learning agent can be divided into three conceptual elements as shown in Fig.02.

Learning element	is responsible for making improvements. The core of the learning element will be a neural network (NN) which will continuously receive sensors data from the agent environment, managing them as input features. This NN has already been trained by using previous operation and performance data retrieved from the chiller plant management system log files. So the time that the learning element is deployed (NN) is ready to receive new real time inputs (from the sensors) and perform several super-fast simulations (algorithm runs) for different equipment configurations. Every time a new run is performed, the NN will predict a value suitable to quantify and measure the performance and consequently the energy efficiency of the chiller plant. This value will be the performance element.
Performance element	is responsible for selecting external actions. The performance element shall be a key performance indicator (KPI) selected to assess and track the chiller plant energy efficiency. Once the learning element try several simulation runs (see line ‘changes’ on Fig.02) and predict the expected performance (through a KPI value), the performance element of the agent will select that scenario with the maximum expected KPI considering system restrictions, boundary conditions and current status. Following this decision, the performance element will deploy a set of corresponding system configurations (actuators) towards the chiller plant.
Critic	measures how the agent is doing and determines how the performance element should be modified to do better in the future. The learning element uses feedback from the critic. The critic tells the learning element how well the agent is doing with respect to a fixed performance standard. It is a part of the algorithm that will receive sensors data back after the chiller plant has received and applied the proposed set-up by the agent. These received sensors data are the real reflection of an initially artificial system behavior predicted by the learning agent. If the measured real system performance is close to the learning agent predicted KPI within a certain percentage, then the proposed system set-up becomes knowledge and added to the learning agent experience (see line ‘knowledge’ on Fig.02).

The above process is repeated continuously and forms an artificial 24/7 system supervisor which does not only measure and monitors the chiller plant status as any energy management system would do. It is an active AI learning agent which measure the performance, collects experience, learns from fault decisions and operates the chiller plant on the most optimum way.

3. Building the Key Performance Indicator (KPI)

The development and implementation of a suitable KPI for measuring the energy efficiency of a chiller plant is indispensable as the performance element of the learning agent. According to C. Schmidt et al. [2], the development of KPIs is characterized by two parallel processes.

First is the design process which focuses on the creation of new indicators and simultaneously the second is to define the prevailing data souring strategy. This strategy needs to be investigated with regards to the availability, collection and storage of data required to run the KPI system. In our case these are the data that are essential to run the full learning agent and come as a collection from sensors (refer to Fig.01).

Since the KPI purpose is to track the chiller plant performance in terms of energy efficiency we have to leave aside performance indicators for individual equipment in isolation. The performance measure shall be a System-Level key performance indicator which will reflect the energy efficiency on a chiller plant level.

According to J.Mayernik [3], the following definition of a system-level KPI shall be the driver of current development.

System-level KPIs quantify energy performance by measuring the capacity of quality of a service provided by a building system per unit of energy used to provide the service.

So according to the definition above, becomes that the building system is the chiller plant along with the chilled water piping network and the terminal cooling units, energy is the electrical energy required for power supply of chillers, pumps, fans etc. and obviously the objective service is the cooling of the building.

So the selected KPI shall be a Type 1 template, according to C. Schmidt et al. [2] with the following form:

KPI = Energy [ … ] per [ … ],

which describes energy costs, consumption or share related to a specific quantity.

For the chiller plant the KPI can be formed as the produced cooling capacity per unit of electrical power consumed to produce this cooling energy.

The produced Cooling Energy is measurable through a heating/cooling energy meter arrangement along the main supply/return pipe lines of the chiller plant (or by more meters installed onto the individual piping circuits).

The consumed Electrical Energy will be measured using a dedicated power meter onto the HVAC system panel.

It is obvious that an energy efficient chiller plant is the one that produces the more cooling energy for the less electrical energy consumption. So ideally we want to operate the chiller plant for those conditions and configuration which returns us back the highest KPI value.

Next step is the collection of data and structure of the learning agent which will allow the development of the algorithm. Before the collection of data the engineer has to select proper chiller plant operation parameters which in the language of AI-Machine Learning is called Input Features, then the extraction of history data and preparation is called training datasets. Finally the learning element will be a Neural Network.

The implementation methodology and technical details for all these are described next.

4. Features Selection

Through this stage we perform the so called features extraction and parameters selection procedure along with the decision of measured outputs and KPIs (labeled data).

Having studied and fully understand the system architecture next step is to identify those system parameters that mainly affect the operation of it. During this feature engineering phase, we have to select those parameters that are mostly linked with the system and somehow feed its internal known and unknown engineering functions.

So diving back to the chiller system we will define features that describe the plant environment and the equipment operation.

A few proposed features of the system are listed below:

Chiller Plant Equipment

1.      No of chillers running; 2.      No of circuits per chiller running; 3.      Chillers LCWT; 4.      No of primary CHW Pumps running; 5.      No of secondary CHW Pumps running; 6.      Secondary CHW pumps load (mean value %); 7.      No of cooling towers running; 8.      No of CDW Pumps running;

External Ambient Environment

9.      Outside air dry bulb temperature; 10.  Outside air wet bulb temperature;

Cooling Load Demand

11.  System cooling demand (kW) – Measured through energy meters.

All above are obviously easy measurable quantities, through sensors or predicted by calculation formulas based on field sensored input data. Fig.01 represents of how the input features data listed above are collected from the agent ‘environment’ through sensors and feed the learning agent.

5. Data Collection and Vectorization

Input source of data both for the ML algorithm training and further operation and continuous system self-learning development shall be the chiller control system monitoring and log files. Either a local chiller plant PLC system or the central BMS or SCADA system can be used.

The extraction of data is strongly related to the available format of data and the establishment of a link that will allow initially the extraction of the training dataset and secondary the real time data gathering that will feed the final API.

The training, evaluation and testing of the model will be based on static data set, meaning that a representative batch of data will be retrieved, stored and prepared for feature extraction

It is proposed once data are securely retrieved, both the inputs and outputs shall be turned into tensors of float32 data.

5.1 Data Normalization

Before we feed the data into the neural network first need to cast it to float32 and convert it into the 0 – 1 range. This is called data normalization and we want to smooth differences on the number of digits and value differences between the input data.

Heterogeneous data, where for example one feature is in the range of 0-1 and another is in the range 100-200, can trigger large gradient updates that will prevent the NN from converging.

To overcome this issue and prepare the training set, the applied steps are as below:

• Normalize each feature independently to have a mean of 0,
• Normalize each feature independently to have a standard deviation of 1.

Once the full dataset is ready and properly prepared, it will be grouped into training, validation and testing sets. 

6. Machine Learning Model (The Learning Element)

As analyzed above, the chiller performance optimization will have to provide the operator with a set of expected performance KPIs.

These values will have to be easily and instantaneously predicted by the ML algorithm during the real time operation of the system.

The selected ML model can be a regression type neural network. This method is selected as a very effective way to simulate responses of non-linear systems. The chiller plant system includes equipment which is difficult to model and follow the non-linear responses of every system component easily and accurately.

The architecture of the model shall have an input layer, several hidden layers and an output single node layer that will give the predicted value.

NN Layers

The number of hidden layers should initially be equal to two or three in order to avoid overfitting. If the predicted results are not satisfactory the number of layers can increase by one each time to check the accuracy.

The proposed configuration is a small network with two hidden layers, each with 128 units.

The network will end with a single unit, a linear layer, which is a typical setup for scalar regression (a regression where we try to predict a single continuous value).

The loss function will be of a mean squared error type, the square of the difference between the prediction and the target performance KPI.

Also the metric during training will be mean absolute error which is the absolute value of the difference between the predictions and the targets again.

Dataset

The prepared datasets will be used to train the ML algorithm and create a non-linear hidden function that will map all input data to the related outputs. As a supervised machine learning application, the initial training of the NN model will have to use the data set which extracted from the history of the chiller plant as described in section 5 above.

The training data set shall be of a (x_i, y_i) format with:

x_i – the i^th input data points training set of the sensors inputs described in section 4.

y_i – the i^th labeled data point which is the measured chiller plant performance value. This is the real achieved performance of the chiller plant. Based on the history data of consumed electrical energy and produced cooling can easily be calculated and stored by using the KPI of section 3.

By this way we will have a training set (deployed for at least one year of operation) with input data and labeled outputs.

7.  Other Topics to be considered

7.1  Chiller Operation Limits

There are always operation limits or what I call system boundary conditions that we have to consider. These elements are:

• The minimum water flow that we need to have through the chiller in order to be considered operational;
• The minimum pressure on the inlet side of the chiller (low pressure side);
• The maximum and minimum ambient air temperature limits that allow the chiller to operate.
• The agent algorithm shall also consider the above limits which will be modelled as alarm points for safe operation of the equipment and limits of equipment configuration decisions by the learning agent.

7.2 Modelling of the Environment

Important parameters of the external environment that affect the operation and performance of chillers and also have to be included within the algorithm are:

• The ambient air dry bulb temperature in case of air cooled chillers;
• The ambient air wet bulb temperature in case of water cooled chillers;
• The ambient air wet bulb temperature in case of evaporative type air cooled chillers;

The above as well is under development in the form of  code along with all interfaces of data transfer in Python. Hope to be in a position to share that soon. In the meantime as long as I would like to test the performance of the agent on real data, I would be happy to support any authority, end user or developer which has a large chiller plant in operation and would like to share data. Please contact me on moumiadis@gmail.com. 

If you did like it please share it through the social so that more people can have access to it. If you have any questions or would like to discuss any special case, please leave your comments below. I will be happy to answer!

References

[1] Stuart Russell and Peter Norvig, ‘Artificial Intelligence – A Modern Approach’, Pearson Education Limited 3^rd ed. 2016,

[2] Christopher Schmidt et al, ‘Implementing Key Performance Indicators for Energy Efficiency in Manufacturing’, 49^th CIRP Conference on Manufacturing Systems (CIRP-CMS 2016),

[3] John Mayernik, ‘System-Level Key Performance Indicators’, ASHRAE High Performing Buildings, Fall 2019, page 31-33.

Abbreviations

CWST – Chilled water supply temperature

CWRT – Chilled water return temperature

CWP – Chilled water pump

CDWP – Condensing water pump

DBT – Dry bulb temperature

WBT – Wet bulb temperature

API – Application Programming Interface