Difference between revisions of "Projects:2019s2-25202 Electricity Consumers' Greenness: A Measure to Quantify Electricity Consumers' Sustainability and Efficiency"

From Projects
Jump to: navigation, search
(→‎Artificial Neural Network (Data Driven model))
(→‎Scenario 2: Consumer Interactions - High)
 
(53 intermediate revisions by the same user not shown)
Line 2: Line 2:
 
[[Category:Final Year Projects]]
 
[[Category:Final Year Projects]]
 
[[Category: 2019s2|25202]]
 
[[Category: 2019s2|25202]]
−
Generation and demand balance is a critical job for the grid operator. This balance is being affected by the increasing number of frequency events on the grid which in turn require fast balancing resources. However, existing balancing resources, such as synchronous generators, are retiring for many economic and environmental reasons. Adding more intermittent renewable generation makes the real-time operation a lot more difficult compared to traditional grid, also requiring higher amount of balancing resources. Therefore, this project focuses on introducing a better way of energy consumption by quantifying the inefficiencies in the consumer behaviour.
+
 
 +
== Abstract ==
 +
Supply and demand balancing resources are retiring due to the ever-increasing addition of intermittent generation on the power system. Vast majority of the demand side consists of residential consumers and is affected by how they interact with their electrical appliances. Major appliances in a household, such as a refrigerator, are used at peak hours during the day. Trying to change consumer behaviour without challenging their comfort level becomes very crucial as it could reduce carbon emissions and also assist in cheaper operation of the power system as a whole.
 +
 +
Electricity Consumers’ Greenness is a proposed method for quantifying and pinpointing the inefficiency in the consumer behaviour. This involves the aggregation of a greenness factor for household appliances. Different design and modelling approaches have been undertaken to create a Greenness map for a household. Therefore, this project focuses on introducing a better way of energy consumption by quantifying the inefficiencies in the consumer behaviour.
 +
 
 
== Introduction ==
 
== Introduction ==
 
Quantifying the energy-efficiency of a building using electricity consumption data has been a research question for many years. These efforts were based on identifying the source of inefficiencies in the building structure or the device itself. However, even in a highly efficient building, consumers with their unique behaviours can degrade the efficiency of energy usage. One example is using an appliance at on-peak hours or setting the HVAC setpoint to an unreasonable level during the same time. As the Internet-of-Things (IoT) is becoming widespread, more information about individual appliance's consumption is available. Therefore, it is time to pinpoint the inefficiency in the consumers' behaviours. In this project, we intend to introduce the "Greenness" factor for a household by first designing a "Greenness" map for every appliance in a household based on their physical and operational limitations, and relevant exogenous factors such as ambient temperature, type of day, the hour of the day, season, etc. Then, we quantify the "Greenness" of a household electricity consumption according to the data for every appliance. In the next step, we develop an aggregate "Greenness" map for every day of every year of the historical data to represent the overall efficiency of the consumer's behaviour in a household. Finally, we convert the aggregate "Greenness" map to a single "Greenness" factor for the household.
 
Quantifying the energy-efficiency of a building using electricity consumption data has been a research question for many years. These efforts were based on identifying the source of inefficiencies in the building structure or the device itself. However, even in a highly efficient building, consumers with their unique behaviours can degrade the efficiency of energy usage. One example is using an appliance at on-peak hours or setting the HVAC setpoint to an unreasonable level during the same time. As the Internet-of-Things (IoT) is becoming widespread, more information about individual appliance's consumption is available. Therefore, it is time to pinpoint the inefficiency in the consumers' behaviours. In this project, we intend to introduce the "Greenness" factor for a household by first designing a "Greenness" map for every appliance in a household based on their physical and operational limitations, and relevant exogenous factors such as ambient temperature, type of day, the hour of the day, season, etc. Then, we quantify the "Greenness" of a household electricity consumption according to the data for every appliance. In the next step, we develop an aggregate "Greenness" map for every day of every year of the historical data to represent the overall efficiency of the consumer's behaviour in a household. Finally, we convert the aggregate "Greenness" map to a single "Greenness" factor for the household.
Line 68: Line 73:
  
  
−
 
−
=== Topic 1 ===
 
  
 
== Methodology ==
 
== Methodology ==
Line 75: Line 78:
  
 
=== Data Acquisition ===
 
=== Data Acquisition ===
 +
 +
Data acquisition is vital to achieve our project objectives. PECAN Street’s online database, Dataport, helps us gain access to approximately 6 years’ of 1-minute interval historic data of electricity use in households. These dynamic datasets are available in .csv format which assist us in easy analysis of consumption data for cyclic and noncyclic household appliances.
 +
 +
A PECAN Street data set has been used to analyse the behaviour of two major household appliances. shows the power consumption (in kW) for a refrigerator (cyclic) and oven (noncyclic) over the course of one day i.e. 00:00 to 23:20. It can be clearly observed that refrigerators exhibit power consumption behaviour in the form of a duty cycle. The on and off transitions from a reference power of approximately 0.1 kW are due to the consumer interaction with the refrigerator plus other factors such as room temperature. However, the operation of an oven is seen as four distinguishable peaks during the day. This is due to the unpredictable consumer interaction with ovens, proving the noncyclic behaviour of the appliance.
 +
 +
PECAN Street Data consists of reliable real-time datasets to analyse electrical consumption of household appliances. However, our main assumption to calculate the base case is that consumer interaction with the appliances is non-existent in certain hours of the day. There is typically little to no interaction of consumers with appliances after midnight to early hours in the day. Our base case is to be derived using the real-time data of the refrigerator during the early hours of the day. This requires the need for a design project which would help us accumulate real power readings in an experimental household.
 +
 
=== Design Approach ===
 
=== Design Approach ===
 +
 +
[[File:Design Experiment for Base Case.png|thumb|center|Figure 2: Design Experiment for deriving Base Case]]
 +
 +
The design approach of our project involves use of an energy monitor, temperature data logger and a timer. The energy monitor is collecting the electrical consumption data from the refrigerator during the day. The use of a timer is assisting us in noting down the duration of interaction at a daily basis and help us analyze the power consumption over a period of 30 days. The temperature data logger is being used to log indoor temperature of the room where the refrigerator is located. This temperature would be used as an exogenous factor input in our modelling approach. Once collected and analysed, this real-time data would then be compared against the derived base case to help us calculate the Greenness Factor of the appliance.
 +
 +
The details of the refrigerator that we have selected for our experiment are tabulated in the table below:
 +
 +
 +
{| class="wikitable"
 +
|-
 +
| Manufacturer || Westinghouse
 +
|-
 +
| Model No.|| WTM3900WBR
 +
|-
 +
| Type || Top Mount
 +
|-
 +
| Rated Gross Volume || 390 L
 +
|-
 +
| Voltage Rating || 220-240 V
 +
|-
 +
| Current Rating  || 0.64 A
 +
|-
 +
| Dimensions || Height:1720 mm
 +
 +
Width:703 mm
 +
 +
Depth:647mm
 +
|-
 +
|}
 +
 +
This refrigerator is a top mount type which means it would be convenient to record the number of consumer interactions and setup the experiment for monitoring the power consumption data.
 +
 +
 +
[[File:Engage Hub.jpg|thumb|Figure 3: Efergy Engage Hub –Single Phase Model, digital photograph, Reduction Revolution, accessed 17 September 2019, [24] https://reductionrevolution.com.au/products/efergy-engage-hub-online-energy-monitor-regular]]
 +
 +
Figure 2 hows the energy meter that we are using for our project. This is an online household power monitor which consists of the following major components:
 +
 +
• Engage Hub
 +
 +
• Transmitter
 +
 +
• CT Sensor Clamp
 +
 +
The wireless hub receives data via the transmitter and sensor clamp attached to the appliance. In comparison to other energy meters available in the market, the Efergy Engage Hub kit has the following advantages:
 +
 +
● Energy Demand - shows a detailed daily load profile (chart) or your usage over time.
 +
 +
● History Usage - shows historical consumption over hours, days, weeks, or months.
 +
 +
● Monthly data – available in csv format; easy analysis of data
 +
 +
[[File:Temp Data logger.jpg|thumb|left|Figure 4: Multi-use USB Temperature Data Logger - IC-RC-5 Single Phase Model, digital photograph, Instrument Choice, accessed 01 June 2020 [25]]]
 +
 +
Figure 3 shows the temperature data logger that we have selected to log indoor temperature. This is a plug-and-play data logger has a 0.1 C resolution and records real-time temperature for up to 32000 data points i.e. approximately equal to 22 days of 1-minute interval data. This means that the data would have to be exported fortnightly to free up memory for a 30 days analysis.
 +
 
=== Modelling approach ===
 
=== Modelling approach ===
  
Line 94: Line 159:
 
|}
 
|}
  
−
shows a summary of the modelling approaches that have been considered for our project. These are data-driven and physics-based respectively.  
+
Table 2 shows a summary of the modelling approaches that have been considered for our project. These are data-driven and physics-based respectively.  
  
 
Data-driven systems treat the system being modelled like a ‘black box’ – a set of training inputs and desired output are defined without any knowledge of the internal workings. This is an accurate approach for non-linear systems as it requires minimum knowledge of the physical aspects of the system. However, large datasets and parameter selections is a tedious and time-taking task in data-driven systems. Data-driven systems involve training through a trial-and-error process to learn and be able to predict from the defined set of data.  
 
Data-driven systems treat the system being modelled like a ‘black box’ – a set of training inputs and desired output are defined without any knowledge of the internal workings. This is an accurate approach for non-linear systems as it requires minimum knowledge of the physical aspects of the system. However, large datasets and parameter selections is a tedious and time-taking task in data-driven systems. Data-driven systems involve training through a trial-and-error process to learn and be able to predict from the defined set of data.  
Line 106: Line 171:
 
The inputs are fed into the input layer which is often a single layer. Larger number of hidden layers would increase the processing time of the neural network to produce an output. The hidden layers contain neurons and perform transformation of the non-linear inputs entered. The output layer is connected to the hidden layer and displays the output of the model.
 
The inputs are fed into the input layer which is often a single layer. Larger number of hidden layers would increase the processing time of the neural network to produce an output. The hidden layers contain neurons and perform transformation of the non-linear inputs entered. The output layer is connected to the hidden layer and displays the output of the model.
  
−
== Results ==
+
== Implementation and Testing ==
 +
=== Experimental setup in a household ===
 +
 
 +
A design experiment was necessary to be implemented in an experimental household to acquire real-time power consumption data and indoor temperature of the room the refrigerator was located.
 +
The CT sensor clamp (provided with the Efergy Engage Hub) has been manufactured to work on the main supply cable behind the switchboard rather the appliance’s power cord. Initial testing of the design experiment recorded no electrical data on our energy meter as the sensor clamp was not able to sense any alternating current. To resolve this issue, two possible approaches have been considered:
 +
1. Removing the outer sheath of the appliance power cord to place sensor clamp onto the live cable
 +
2. Use of an experimental measurement box
 +
 
 +
 
 +
[[File:Experimental setup.jpg|thumb|center|Figure 5: Experimental Setup]]
 +
 
 +
 
 +
 
 +
The measurement box has two loops which can be used with our CT sensor clamp. The electrical appliance can be connected at one end of the extension (grey connector on the left) and the CT sensor clamp onto the thinner loop (one turn of wire). This box would serve the purpose of an extension lead.
 +
The measurement box would be more appropriate to consider from both an accuracy and safety point of view. The advantage of using this measurement box is the availability of the ten-turn loop in case the sensor clamp for better accuracy.
 +
 
 +
=== Neural Network implementation ===
 +
 
 +
Our inputs for this research are power consumption of the refrigerator and the indoor temperature of the room the refrigerator is located. The indoor temperature is one of the exogenous factors being considered in our experiment. Since these inputs are non-linear, we need a nonlinear dynamic network which takes exogenous inputs.
 +
 
 +
The NARX (Nonlinear autoregressive with external input) model is being implemented in our project that utilizes observations from a past time series to predict the future values of that time series. This network uses tapped delay lines to store the previous values of the input sequences. It is important to note that the output of the NARX network is being fed back to the input of the network through delays.
 +
The feedback loop can be open loop for efficient training of the NARX network. Since we define the output of the network, an open-loop architecture is ideal for training. The number of neurons and delays can be set accordingly depending on the training performance of the network.
 +
 
 +
We are using the default number of neurons for our network i.e. 10 and the number of delays has been set to 30 instead of the default number 2. Figure 5 illustrates how the input data is selected for training. There are total 1440 minutes in a day. By defining a delay of 30, the network selects the first 30 input values for training and predicts the 31st value i.e. the target value.
 +
 
 +
[[File:Training ANN.png|thumb|center|Figure 6: Selection of data for training Neural Network ]]
 +
 
 +
[[File:Error Histogram.jpg|thumb|center|Figure 7: Error (close to zero) displays the performance of trained neural network]]
 +
 
 +
=== Dynamic Base Case ===
 +
 
 +
After training the neural network, we obtain a base case that would be our point of reference for future predictions. The target outputs are fed back into the network as binary inputs denoting ON/OFF states of the refrigerator. Since the compressor of the refrigerator causes uncertain changes in the electrical consumption data, the binary inputs simplify the data for further analysis.
 +
 
 +
We need a network that can predict the consumer behaviour accurately. The consumer interactions can vary from short to long intervals during the day. In order to quantify that, the corresponding binary inputs and outputs of the power consumption data are considered. The trained neural network predicts the ON/OFF state of the refrigerator inclusive of the consumer interactions. This gives us our dynamic base case.
 +
The dynamic base case represents the new, predicted base case. The indoor temperature data is inputted in the trained network to predict the ON/OFF state of the refrigerator due to that indoor temperature. However, we see differences between our dynamic base case and the actual data when compared. These differences signify the consumer interaction with the appliance.
 +
 
 +
We need to incorporate a method which would help us signify the differences. It is important not to penalise the consumer where they have not interacted with the device, but the differences would suggest otherwise. The actual base case would have an ON state for the refrigerator when the temperature is above a certain limit. This would cause the dynamic base case to predict in a similar fashion. However, in reality the consumer may have interacted with the refrigerator a few minutes back in time and an interaction would be showing up due to the compressor switching on later. This would result in anomalies leading to unfair penalties being imposed on the consumer. Therefore, we need to account for these interactions which may accumulate to give significant differences throughout the day.
 +
 
 +
=== Concept of Accumulated Credit ===
 +
 
 +
Credit value acts as a counter to accommodate the shift in normal power usage due to an interaction by the consumer. Credit value works in the following manner:
 +
 +
• If Dynamic Base Case = 1 and Actual values = 0, Credit=Credit+1
 +
 
 +
• If Dynamic Base Case = 0 and Actual values = 1, Credit=Credit-1
 +
 
 +
• If Dynamic Base Case = 1 and Actual values = 1, Credit= no change
 +
 
 +
Credit value starts counting positively when the dynamic base case is on and the compressor of the fridge is off. The value of credit keeps adding until the compressor is finally switched on and matches the dynamic base case. The credit value remains unchanged when both the dynamic base case and actual values of the original base match. If the fridge is on during a high demand period i.e. high electricity price, credit starts subtracting until and unless it matches the dynamic base case again. This ensures that the consumer is only penalised when they are interacting with the fridge during peak demand hours.
 +
 
 +
The dynamic base case would always be lower than the actual values because actual values include consumer interaction.
 +
 
 +
=== Greenness Factor of a refrigerator in a household ===
 +
 
 +
After obtaining a predictive NARX model with indoor temperature as the exogenous factor, the Greenness factor of a refrigerator can now be predicted. We need a way of relating the accumulated credit values to the electricity grid and accurately signify the greenness factor of the refrigerator. Electricity spot prices are the ideal way of representing the greenness factor in an acceptable, clear layout.
 +
 
 +
The consumer interaction can now be predicted against the electricity spot prices. An average of the electricity prices during the day can be plotted and the consumer interactions predicted depending on the spot price. If the consumer interaction is higher than the average price, that would be inefficient towards the grid. If the consumer interaction is lower than the average price, that would be efficient towards the grid.
 +
 
 +
== Results and Discussion ==
 +
 
 +
=== Base Case ===
 +
 
 +
The typical power consumption of a domestic refrigerator lies between 1-2 kWh which can be seen on the graph.
 +
However, when this data was analysed closely, some of the data points were either missing or distorted. Figure 18 illustrates the distorted data for two days. This may have been caused due to the electromagnetic radiation from the refrigerator compressor.
 +
 
 +
We required clean, uninterrupted data to create our base case. It is important to point out that the recorded data via the energy meter contained vertical spikes and surges through the entirety of each day. Although these uncertain peaks seem inaccurate, they were safe to be neglected as our main analysis is based on the time of interaction i.e. the horizontal width of the ON state.
 +
 
 +
The data was screened for anomalies and cleaned to obtain accurate datasets for further analysis.
 +
The base case was created using the hours where consumer interaction was negligible. A ‘low’ data set was used for training the NARX network. ‘Low’ here means that the consumer had little to no interaction during the entire day in real-time.
 +
 +
Since the base case was created in the hidden layers of the neural network, the respective datasets taken as inputs for the base case derivation have been discussed. Figure 19 shows the power consumption data of the refrigerator in the early hours of the day. Indoor temperature was our exogenous factor to be considered for training our neural network.
 +
 
 +
=== Consumer Interaction ===
 +
 
 +
[[File:Consumer interact.png|thumb|right|Table 3: Consumer Interaction with the refrigerator]]
 +
 
 +
Four datasets were created to model the types of consumer interaction in a household as shown in Table 3. This involved the interaction of one of the project members with the refrigerator to create different types of consumer interaction. Short interactions involve taking something out of the fridge and the opening/closing duration is short. These interactions lie in the low to medium dataset category. Longer interactions include keeping the fridge open for longer duration like for example replenishing the fridge with groceries. These lie in the high category. The Very High category shows a very rare situation of interactions where each interaction is approximately 2 minutes.
 +
 
 +
=== Generation of Greenness Factor ===
 +
 
 +
The Greenness factor has been generated for two scenarios:
 +
 
 +
• Scenario 1: Morning-Midday, Medium interactions
 +
 
 +
• Scenario 2: Early Evening, High interactions
 +
 
 +
Scenario 1 predicts the situation when the consumer might be outside during the day and might have family members interacting with the refrigerator. This would account for medium interactions during the morning-midday time. Scenario 2 predicts that the consumer has just arrived home after close of business (around 17:00 hrs) and interacting with the refrigerator. This would involve high interactions with the refrigerator due to the frequent opening and closing of the refrigerator in comparison to other hours of the day.
 +
 
 +
 
 +
==== Scenario 1: Consumer Interactions - Medium ====
 +
 
 +
The number of consumer interactions are shown by the red lines on the graph and lie in the morning to midday i.e. 07:30 to 13:30 hours. The average spot price of the whole day has been shown by the horizontal broken line that lies approximately around $34/MWh. The consumer interactions have been predicted large in number but since these instances lie under the average spot price, the consumer is green
 +
 
 +
[[File:Med interact.png|thumb|center|Figure 8: Greenness Factor – Medium number of interactions]]
 +
 
 +
==== Scenario 2: Consumer Interactions - High ====
 +
 
 +
The number of consumer interactions are shown by the red lines on the graph and lie in the morning to midday hours. The average spot price of the whole day has been shown by the horizontal broken line that lies approximately around $20/MWh. Consumer interactions are predicted to lie throughout the day, most lying around the 1000 minutes range. This is equivalent to a time of 16:00-17:00 hours which is usually close of business and consumers are interacting more with their refrigerator. The key point to note here is that some of the consumer interactions are crossing the average spot price. This shows that the consumer is red in these instances while staying green rest of the day.
 +
 
 +
[[File:Higher interact.png|thumb|center|Figure 9: Greenness Factor – High number of interactions]]
  
 
== Conclusion ==
 
== Conclusion ==
 +
 +
There are many key points that can be concluded from this project. A typical household consists of two different types of major appliances – cyclic and noncyclic. This requires the need for modelling a base case operation for cyclic appliances so they can be compared and evaluated against a greenness factor. The Greenness factor would then allow us to quantify the differences in the electricity consumption profiles.
 +
 +
Secondly, it is vital to screen data acquired from the energy for distortions. Distortions might be present in the data in the form of unwanted surges or spikes. This may be arising due to the presence of electromagnetic radiation from surrounding appliances or the refrigerator’s own compressor.
 +
 +
We learned that a data-driven approach works better for nonlinear loads. Since our input datasets were in a time series, we learned about nonlinear autoregressive networks with external input and how they are ideal for a time series prediction. It was interesting to experiment with different number of neurons and delays to see how the neural networks efficiently trained the network to predict more accurately. However, efficient training of the network also required larger datasets and proved to be time-consuming.
 +
 +
Although it was difficult to establish a relation between the greenness factor inclusive of the accumulated credit with the electricity spot prices, the generated greenness factor helped in understanding the impact of exogenous factors on the consumer interaction.
 +
 +
All in all, it is safe to conclude that Greenness factor is an efficient electricity solution that can transform the way consumers interact with our household appliances. The assumption that consumer interactions are inefficient towards the electrical grid is accurate according to our results. However, our predictive model can be trained more accurately by considering other exogenous factors as well such as type of room/day, outside temperature, etc. to generate a more reliable greenness factor.
  
 
== References ==
 
== References ==
Line 119: Line 293:
  
 
[4] Kussel, G, Frondel, M, Switching on electricity demand response: Evidence for German households, EconPapers, 2016
 
[4] Kussel, G, Frondel, M, Switching on electricity demand response: Evidence for German households, EconPapers, 2016
 +
 +
[5] I. J. U. P. Stadler, "Power grid balancing of energy systems with high renewable energy penetration by demand response," vol. 16, no. 2, pp. 90-98, 2008.
 +
 +
[6] P. Tielens and D. Van Hertem, "Grid inertia and frequency control in power systems with high penetration of renewables," in Young Researchers Symposium in Electrical Power Engineering, Date: 2012/04/16-2012/04/17, Location: Delft, The Netherlands, 2012.
 +
 +
[7] K. J. E. E. Gram-Hanssen, "Efficient technologies or user behaviour, which is the more important when reducing households’ energy consumption?," vol. 6, no. 3, pp. 447-457, 2013.
 +
 +
[8] J.-M. Cayla, N. Maizi, and C. J. E. p. Marchand, "The role of income in energy consumption behaviour: Evidence from French households data," vol. 39, no. 12, pp. 7874-7883, 2011.
 +
 +
[9] M. A. Khan and A. J. O. E. R. Qayyum, "The demand for electricity in Pakistan," vol. 33, no. 1, pp. 70-96, 2009.
 +
 +
[10] G. J. E. E. Hondroyiannis, "Estimating residential demand for electricity in Greece," vol. 26, no. 3, pp. 319-334, 2004.
 +
 +
[11] C. Wolfram, O. Shelef, and P. J. J. o. E. P. Gertler, "How will energy demand develop in the developing world?," vol. 26, no. 1, pp. 119-38, 2012.
 +
 +
[12] M. Muratori, M. C. Roberts, R. Sioshansi, V. Marano, and G. J. A. E. Rizzoni, "A highly resolved modeling technique to simulate residential power demand," vol. 107, pp. 465-473, 2013.
 +
 +
[13] A. Foucquier, S. Robert, F. Suard, L. Stéphan, A. J. R. Jay, and S. E. Reviews, "State of the art in building modelling and energy performances prediction: A review," vol. 23, pp. 272-288, 2013.
 +
 +
[14] C. D. Himmel and G. S. J. I. T. o. s. m. May, "Advantages of plasma etch modeling using neural networks over statistical techniques," vol. 6, no. 2, pp. 103-111, 1993.
 +
 +
[15] L. G. Swan, V. I. J. R. Ugursal, and s. e. reviews, "Modeling of end-use energy consumption in the residential sector: A review of modeling techniques," vol. 13, no. 8, pp. 1819-1835, 2009.
 +
 +
[16] D. J. Livingstone, D. T. Manallack, and I. V. J. J. o. c.-a. m. d. Tetko, "Data modelling with neural networks: advantages and limitations," vol. 11, no. 2, pp. 135-142, 1997.
 +
 +
[17] F. Ziel, R. Steinert, and S. J. E. E. Husmann, "Efficient modeling and forecasting of electricity spot prices," vol. 47, pp. 98-111, 2015.
 +
 +
[18] P. S. database. (2019). Available: https://www.pecanstreet.org/dataport/
 +
 +
[19] F. Burden and D. Winkler, "Bayesian regularization of neural networks," in Artificial neural networks: Springer, 2008, pp. 23-42.
 +
 +
[20] AEMO. ACCESS MARKET PORTALS. Available: https://aemo.com.au/energy-systems/electricity/national-electricity-market-nem/data-nem/data-dashboard-nem

Latest revision as of 11:59, 9 June 2020


Abstract

Supply and demand balancing resources are retiring due to the ever-increasing addition of intermittent generation on the power system. Vast majority of the demand side consists of residential consumers and is affected by how they interact with their electrical appliances. Major appliances in a household, such as a refrigerator, are used at peak hours during the day. Trying to change consumer behaviour without challenging their comfort level becomes very crucial as it could reduce carbon emissions and also assist in cheaper operation of the power system as a whole.

Electricity Consumers’ Greenness is a proposed method for quantifying and pinpointing the inefficiency in the consumer behaviour. This involves the aggregation of a greenness factor for household appliances. Different design and modelling approaches have been undertaken to create a Greenness map for a household. Therefore, this project focuses on introducing a better way of energy consumption by quantifying the inefficiencies in the consumer behaviour.

Introduction

Quantifying the energy-efficiency of a building using electricity consumption data has been a research question for many years. These efforts were based on identifying the source of inefficiencies in the building structure or the device itself. However, even in a highly efficient building, consumers with their unique behaviours can degrade the efficiency of energy usage. One example is using an appliance at on-peak hours or setting the HVAC setpoint to an unreasonable level during the same time. As the Internet-of-Things (IoT) is becoming widespread, more information about individual appliance's consumption is available. Therefore, it is time to pinpoint the inefficiency in the consumers' behaviours. In this project, we intend to introduce the "Greenness" factor for a household by first designing a "Greenness" map for every appliance in a household based on their physical and operational limitations, and relevant exogenous factors such as ambient temperature, type of day, the hour of the day, season, etc. Then, we quantify the "Greenness" of a household electricity consumption according to the data for every appliance. In the next step, we develop an aggregate "Greenness" map for every day of every year of the historical data to represent the overall efficiency of the consumer's behaviour in a household. Finally, we convert the aggregate "Greenness" map to a single "Greenness" factor for the household.

Project team

Project students

  • Usman Khan
  • Farrukh Ahsan
  • Bo Yin

Supervisors

  • Ali Pourmousavi Kani

Advisors

  • Associate Professor Nesimi Ertugrul

Objectives

1. Design a "Greenness" map for every major appliance in a household based on their physical and operational limitations.

2. Account for the exogenous factors such as ambient temperature, type of day, the hour of the day, season, etc. when needed.

3. Aggregating the "Greenness" map for every day of every year of the historical data to represent the overall efficiency of the consumers’ behaviour in a household.

Background

There has been no previous work done on the estimation of Electricity Consumers’ Greenness. However, there are other areas of research that can be related to Consumers’ Greenness. The following table is a summary of some key research focus areas and their impact on consumer behaviour.

Table 1: Pros and Cons of Research Focus Areas related to Consumers' Greenness
Focus of Study Pros Cons
Residential Energy Efficiency Ratings • Highlight ability of efficient households to satisfy consumers • Not influential due to lack of consumers’ engagement in relation to sustainability certification and implementation

• Static calculation discarding long-term effects


Energy Rating Labels on Appliances • Communicate useful energy information • No significant behavioral change
Energy Audit • Energy based renovation is driven by householder perception of comfort and acceptable outlay on energy bills • Complex outlay

• High payback periods

Demand Response • Flexible price design

• No compromise of consumer privacy

• Electricity demand of uninformed households' price-inelastic

• No real-time power

Non-Intrusive Load Monitoring • Easy to monitor energy consumption

• Flexible options for grid operator

• Insufficient point of references for applicability


Most of the research carried out in the field of engineering involves assumption of certain conditions or quantities being constant in order to observe the key parameters being considered. This is usually carried out while formulating equations or cases that serve as the point of reference for other data parameters to be compared and tested against. Similarly, Electricity Consumers’ Greenness requires the formulation of a point of reference against which normal behaviour of major appliances can be compared. This point of reference can be defined as the base case. The base case involves monitoring the consumption data of major appliances in a household during the time in which consumers have negligible interaction with the appliances (e.g. 12 AM to 5 AM). We are assuming that consumers would not be interacting with the household appliances after midnight to early hours of the day. Our main focus here is to consider household appliances that follow a repetitive set cycle throughout the day. Appliances such as heat pumps and refrigerators follow this cyclic pattern of electricity consumption, therefore known as cyclic appliances. We have chosen refrigerators as the key appliance in our project because they consume electricity in an ‘ON/OFF’ duty cycle rather than one continuous cycle of operation. The base case can be formulated once we obtain the electrical consumption data of a refrigerator during hours without consumer interaction. Next step would involve the observation of real consumption data with consumer interaction during the day (peak hours).


Figure 1: Power vs time graph for Base Case/Real Data comparison


Figure 1 is a clear representation of what we primarily intend to achieve in this project. The base case shows a cyclic behaviour of power consumption against time. The real data i.e. the power consumed by the appliance inclusive of consumer behaviour is shown to have differences as compared to the base case. It is important to note that these differences occur for different time widths at the same power level. These differences signify the inefficiencies of the appliance, which can then be quantified further to give us a Greenness factor of the appliance during the day.


Methodology

Data Acquisition

Data acquisition is vital to achieve our project objectives. PECAN Street’s online database, Dataport, helps us gain access to approximately 6 years’ of 1-minute interval historic data of electricity use in households. These dynamic datasets are available in .csv format which assist us in easy analysis of consumption data for cyclic and noncyclic household appliances.

A PECAN Street data set has been used to analyse the behaviour of two major household appliances. shows the power consumption (in kW) for a refrigerator (cyclic) and oven (noncyclic) over the course of one day i.e. 00:00 to 23:20. It can be clearly observed that refrigerators exhibit power consumption behaviour in the form of a duty cycle. The on and off transitions from a reference power of approximately 0.1 kW are due to the consumer interaction with the refrigerator plus other factors such as room temperature. However, the operation of an oven is seen as four distinguishable peaks during the day. This is due to the unpredictable consumer interaction with ovens, proving the noncyclic behaviour of the appliance.

PECAN Street Data consists of reliable real-time datasets to analyse electrical consumption of household appliances. However, our main assumption to calculate the base case is that consumer interaction with the appliances is non-existent in certain hours of the day. There is typically little to no interaction of consumers with appliances after midnight to early hours in the day. Our base case is to be derived using the real-time data of the refrigerator during the early hours of the day. This requires the need for a design project which would help us accumulate real power readings in an experimental household.

Design Approach

Figure 2: Design Experiment for deriving Base Case

The design approach of our project involves use of an energy monitor, temperature data logger and a timer. The energy monitor is collecting the electrical consumption data from the refrigerator during the day. The use of a timer is assisting us in noting down the duration of interaction at a daily basis and help us analyze the power consumption over a period of 30 days. The temperature data logger is being used to log indoor temperature of the room where the refrigerator is located. This temperature would be used as an exogenous factor input in our modelling approach. Once collected and analysed, this real-time data would then be compared against the derived base case to help us calculate the Greenness Factor of the appliance.

The details of the refrigerator that we have selected for our experiment are tabulated in the table below:


Manufacturer Westinghouse
Model No. WTM3900WBR
Type Top Mount
Rated Gross Volume 390 L
Voltage Rating 220-240 V
Current Rating 0.64 A
Dimensions Height:1720 mm

Width:703 mm

Depth:647mm

This refrigerator is a top mount type which means it would be convenient to record the number of consumer interactions and setup the experiment for monitoring the power consumption data.


Figure 3: Efergy Engage Hub –Single Phase Model, digital photograph, Reduction Revolution, accessed 17 September 2019, [24] https://reductionrevolution.com.au/products/efergy-engage-hub-online-energy-monitor-regular

Figure 2 hows the energy meter that we are using for our project. This is an online household power monitor which consists of the following major components:

• Engage Hub

• Transmitter

• CT Sensor Clamp

The wireless hub receives data via the transmitter and sensor clamp attached to the appliance. In comparison to other energy meters available in the market, the Efergy Engage Hub kit has the following advantages:

● Energy Demand - shows a detailed daily load profile (chart) or your usage over time.

● History Usage - shows historical consumption over hours, days, weeks, or months.

● Monthly data – available in csv format; easy analysis of data

Figure 4: Multi-use USB Temperature Data Logger - IC-RC-5 Single Phase Model, digital photograph, Instrument Choice, accessed 01 June 2020 [25]

Figure 3 shows the temperature data logger that we have selected to log indoor temperature. This is a plug-and-play data logger has a 0.1 C resolution and records real-time temperature for up to 32000 data points i.e. approximately equal to 22 days of 1-minute interval data. This means that the data would have to be exported fortnightly to free up memory for a 30 days analysis.

Modelling approach

Different modelling approaches

Table 2: Summary of Data-driven vs Physics-based approaches
Data-driven Physics-Based
Accurate approach for nonlinear system networks System based on state variables represented by linearized state equations
Requires Less knowledge of physical system Requires extensive knowledge of factors/inputs
Large data is needed to develop an accurate model Assumptions in linearizing model can lead to errors
Training is time consuming Actual behaviour might be different from linearized model

Table 2 shows a summary of the modelling approaches that have been considered for our project. These are data-driven and physics-based respectively.

Data-driven systems treat the system being modelled like a ‘black box’ – a set of training inputs and desired output are defined without any knowledge of the internal workings. This is an accurate approach for non-linear systems as it requires minimum knowledge of the physical aspects of the system. However, large datasets and parameter selections is a tedious and time-taking task in data-driven systems. Data-driven systems involve training through a trial-and-error process to learn and be able to predict from the defined set of data. Physics-based systems employ a mathematical approach, consisting of physical variables represented by linearized state equations. This requires extensive knowledge of the physical system before defining the set of state variables. The linearized model in a physics-based system is based on assumption of some variables which linearize the non-linear systems. This introduces errors in a non-linear, dynamic system. Our project involves numerous unpredictable inputs, such as consumer behavior and exogenous factors, and a targeted (desired) output. A data-driven model best suits our modelling approach to obtain the base case as we prioritize a learning, iterative process in comparison to a linearized, defined model. It is important to develop a sustainable system which learns from data sets and accurately predicts possible situations based on intuition.

Artificial Neural Network (Data Driven model)

Amongst numerous deep learning toolboxes, we have chosen the Artificial Neural Networks (ANN) to serve as our data-driven model. ANN functions like a human mind, consisting of interconnected neurons with a capability to learn from a training data set and predict results accordingly.

The inputs are fed into the input layer which is often a single layer. Larger number of hidden layers would increase the processing time of the neural network to produce an output. The hidden layers contain neurons and perform transformation of the non-linear inputs entered. The output layer is connected to the hidden layer and displays the output of the model.

Implementation and Testing

Experimental setup in a household

A design experiment was necessary to be implemented in an experimental household to acquire real-time power consumption data and indoor temperature of the room the refrigerator was located. The CT sensor clamp (provided with the Efergy Engage Hub) has been manufactured to work on the main supply cable behind the switchboard rather the appliance’s power cord. Initial testing of the design experiment recorded no electrical data on our energy meter as the sensor clamp was not able to sense any alternating current. To resolve this issue, two possible approaches have been considered: 1. Removing the outer sheath of the appliance power cord to place sensor clamp onto the live cable 2. Use of an experimental measurement box


Figure 5: Experimental Setup


The measurement box has two loops which can be used with our CT sensor clamp. The electrical appliance can be connected at one end of the extension (grey connector on the left) and the CT sensor clamp onto the thinner loop (one turn of wire). This box would serve the purpose of an extension lead. The measurement box would be more appropriate to consider from both an accuracy and safety point of view. The advantage of using this measurement box is the availability of the ten-turn loop in case the sensor clamp for better accuracy.

Neural Network implementation

Our inputs for this research are power consumption of the refrigerator and the indoor temperature of the room the refrigerator is located. The indoor temperature is one of the exogenous factors being considered in our experiment. Since these inputs are non-linear, we need a nonlinear dynamic network which takes exogenous inputs.

The NARX (Nonlinear autoregressive with external input) model is being implemented in our project that utilizes observations from a past time series to predict the future values of that time series. This network uses tapped delay lines to store the previous values of the input sequences. It is important to note that the output of the NARX network is being fed back to the input of the network through delays. The feedback loop can be open loop for efficient training of the NARX network. Since we define the output of the network, an open-loop architecture is ideal for training. The number of neurons and delays can be set accordingly depending on the training performance of the network.

We are using the default number of neurons for our network i.e. 10 and the number of delays has been set to 30 instead of the default number 2. Figure 5 illustrates how the input data is selected for training. There are total 1440 minutes in a day. By defining a delay of 30, the network selects the first 30 input values for training and predicts the 31st value i.e. the target value.

Figure 6: Selection of data for training Neural Network
Figure 7: Error (close to zero) displays the performance of trained neural network

Dynamic Base Case

After training the neural network, we obtain a base case that would be our point of reference for future predictions. The target outputs are fed back into the network as binary inputs denoting ON/OFF states of the refrigerator. Since the compressor of the refrigerator causes uncertain changes in the electrical consumption data, the binary inputs simplify the data for further analysis.

We need a network that can predict the consumer behaviour accurately. The consumer interactions can vary from short to long intervals during the day. In order to quantify that, the corresponding binary inputs and outputs of the power consumption data are considered. The trained neural network predicts the ON/OFF state of the refrigerator inclusive of the consumer interactions. This gives us our dynamic base case. The dynamic base case represents the new, predicted base case. The indoor temperature data is inputted in the trained network to predict the ON/OFF state of the refrigerator due to that indoor temperature. However, we see differences between our dynamic base case and the actual data when compared. These differences signify the consumer interaction with the appliance.

We need to incorporate a method which would help us signify the differences. It is important not to penalise the consumer where they have not interacted with the device, but the differences would suggest otherwise. The actual base case would have an ON state for the refrigerator when the temperature is above a certain limit. This would cause the dynamic base case to predict in a similar fashion. However, in reality the consumer may have interacted with the refrigerator a few minutes back in time and an interaction would be showing up due to the compressor switching on later. This would result in anomalies leading to unfair penalties being imposed on the consumer. Therefore, we need to account for these interactions which may accumulate to give significant differences throughout the day.

Concept of Accumulated Credit

Credit value acts as a counter to accommodate the shift in normal power usage due to an interaction by the consumer. Credit value works in the following manner:

• If Dynamic Base Case = 1 and Actual values = 0, Credit=Credit+1

• If Dynamic Base Case = 0 and Actual values = 1, Credit=Credit-1

• If Dynamic Base Case = 1 and Actual values = 1, Credit= no change

Credit value starts counting positively when the dynamic base case is on and the compressor of the fridge is off. The value of credit keeps adding until the compressor is finally switched on and matches the dynamic base case. The credit value remains unchanged when both the dynamic base case and actual values of the original base match. If the fridge is on during a high demand period i.e. high electricity price, credit starts subtracting until and unless it matches the dynamic base case again. This ensures that the consumer is only penalised when they are interacting with the fridge during peak demand hours.

The dynamic base case would always be lower than the actual values because actual values include consumer interaction.

Greenness Factor of a refrigerator in a household

After obtaining a predictive NARX model with indoor temperature as the exogenous factor, the Greenness factor of a refrigerator can now be predicted. We need a way of relating the accumulated credit values to the electricity grid and accurately signify the greenness factor of the refrigerator. Electricity spot prices are the ideal way of representing the greenness factor in an acceptable, clear layout.

The consumer interaction can now be predicted against the electricity spot prices. An average of the electricity prices during the day can be plotted and the consumer interactions predicted depending on the spot price. If the consumer interaction is higher than the average price, that would be inefficient towards the grid. If the consumer interaction is lower than the average price, that would be efficient towards the grid.

Results and Discussion

Base Case

The typical power consumption of a domestic refrigerator lies between 1-2 kWh which can be seen on the graph. However, when this data was analysed closely, some of the data points were either missing or distorted. Figure 18 illustrates the distorted data for two days. This may have been caused due to the electromagnetic radiation from the refrigerator compressor.

We required clean, uninterrupted data to create our base case. It is important to point out that the recorded data via the energy meter contained vertical spikes and surges through the entirety of each day. Although these uncertain peaks seem inaccurate, they were safe to be neglected as our main analysis is based on the time of interaction i.e. the horizontal width of the ON state.

The data was screened for anomalies and cleaned to obtain accurate datasets for further analysis. The base case was created using the hours where consumer interaction was negligible. A ‘low’ data set was used for training the NARX network. ‘Low’ here means that the consumer had little to no interaction during the entire day in real-time.

Since the base case was created in the hidden layers of the neural network, the respective datasets taken as inputs for the base case derivation have been discussed. Figure 19 shows the power consumption data of the refrigerator in the early hours of the day. Indoor temperature was our exogenous factor to be considered for training our neural network.

Consumer Interaction

Table 3: Consumer Interaction with the refrigerator

Four datasets were created to model the types of consumer interaction in a household as shown in Table 3. This involved the interaction of one of the project members with the refrigerator to create different types of consumer interaction. Short interactions involve taking something out of the fridge and the opening/closing duration is short. These interactions lie in the low to medium dataset category. Longer interactions include keeping the fridge open for longer duration like for example replenishing the fridge with groceries. These lie in the high category. The Very High category shows a very rare situation of interactions where each interaction is approximately 2 minutes.

Generation of Greenness Factor

The Greenness factor has been generated for two scenarios:

• Scenario 1: Morning-Midday, Medium interactions

• Scenario 2: Early Evening, High interactions

Scenario 1 predicts the situation when the consumer might be outside during the day and might have family members interacting with the refrigerator. This would account for medium interactions during the morning-midday time. Scenario 2 predicts that the consumer has just arrived home after close of business (around 17:00 hrs) and interacting with the refrigerator. This would involve high interactions with the refrigerator due to the frequent opening and closing of the refrigerator in comparison to other hours of the day.


Scenario 1: Consumer Interactions - Medium

The number of consumer interactions are shown by the red lines on the graph and lie in the morning to midday i.e. 07:30 to 13:30 hours. The average spot price of the whole day has been shown by the horizontal broken line that lies approximately around $34/MWh. The consumer interactions have been predicted large in number but since these instances lie under the average spot price, the consumer is green

Figure 8: Greenness Factor – Medium number of interactions

Scenario 2: Consumer Interactions - High

The number of consumer interactions are shown by the red lines on the graph and lie in the morning to midday hours. The average spot price of the whole day has been shown by the horizontal broken line that lies approximately around $20/MWh. Consumer interactions are predicted to lie throughout the day, most lying around the 1000 minutes range. This is equivalent to a time of 16:00-17:00 hours which is usually close of business and consumers are interacting more with their refrigerator. The key point to note here is that some of the consumer interactions are crossing the average spot price. This shows that the consumer is red in these instances while staying green rest of the day.

Figure 9: Greenness Factor – High number of interactions

Conclusion

There are many key points that can be concluded from this project. A typical household consists of two different types of major appliances – cyclic and noncyclic. This requires the need for modelling a base case operation for cyclic appliances so they can be compared and evaluated against a greenness factor. The Greenness factor would then allow us to quantify the differences in the electricity consumption profiles.

Secondly, it is vital to screen data acquired from the energy for distortions. Distortions might be present in the data in the form of unwanted surges or spikes. This may be arising due to the presence of electromagnetic radiation from surrounding appliances or the refrigerator’s own compressor.

We learned that a data-driven approach works better for nonlinear loads. Since our input datasets were in a time series, we learned about nonlinear autoregressive networks with external input and how they are ideal for a time series prediction. It was interesting to experiment with different number of neurons and delays to see how the neural networks efficiently trained the network to predict more accurately. However, efficient training of the network also required larger datasets and proved to be time-consuming.

Although it was difficult to establish a relation between the greenness factor inclusive of the accumulated credit with the electricity spot prices, the generated greenness factor helped in understanding the impact of exogenous factors on the consumer interaction.

All in all, it is safe to conclude that Greenness factor is an efficient electricity solution that can transform the way consumers interact with our household appliances. The assumption that consumer interactions are inefficient towards the electrical grid is accurate according to our results. However, our predictive model can be trained more accurately by considering other exogenous factors as well such as type of room/day, outside temperature, etc. to generate a more reliable greenness factor.

References

[1] McNeill, D, Wilkie, W, Public Policy and Consumer Information: Impact of the New Energy Labels, Journal of Consumer Research, vol. 6, 1979

[2] Murphy, L, The influence of energy audits on the energy efficiency investments of private owner-occupied households in the Netherlands, ScienceDirect, 2014

[3] Source: Pecan Street Inc. Dataport, 2019

[4] Kussel, G, Frondel, M, Switching on electricity demand response: Evidence for German households, EconPapers, 2016

[5] I. J. U. P. Stadler, "Power grid balancing of energy systems with high renewable energy penetration by demand response," vol. 16, no. 2, pp. 90-98, 2008.

[6] P. Tielens and D. Van Hertem, "Grid inertia and frequency control in power systems with high penetration of renewables," in Young Researchers Symposium in Electrical Power Engineering, Date: 2012/04/16-2012/04/17, Location: Delft, The Netherlands, 2012.

[7] K. J. E. E. Gram-Hanssen, "Efficient technologies or user behaviour, which is the more important when reducing households’ energy consumption?," vol. 6, no. 3, pp. 447-457, 2013.

[8] J.-M. Cayla, N. Maizi, and C. J. E. p. Marchand, "The role of income in energy consumption behaviour: Evidence from French households data," vol. 39, no. 12, pp. 7874-7883, 2011.

[9] M. A. Khan and A. J. O. E. R. Qayyum, "The demand for electricity in Pakistan," vol. 33, no. 1, pp. 70-96, 2009.

[10] G. J. E. E. Hondroyiannis, "Estimating residential demand for electricity in Greece," vol. 26, no. 3, pp. 319-334, 2004.

[11] C. Wolfram, O. Shelef, and P. J. J. o. E. P. Gertler, "How will energy demand develop in the developing world?," vol. 26, no. 1, pp. 119-38, 2012.

[12] M. Muratori, M. C. Roberts, R. Sioshansi, V. Marano, and G. J. A. E. Rizzoni, "A highly resolved modeling technique to simulate residential power demand," vol. 107, pp. 465-473, 2013.

[13] A. Foucquier, S. Robert, F. Suard, L. Stéphan, A. J. R. Jay, and S. E. Reviews, "State of the art in building modelling and energy performances prediction: A review," vol. 23, pp. 272-288, 2013.

[14] C. D. Himmel and G. S. J. I. T. o. s. m. May, "Advantages of plasma etch modeling using neural networks over statistical techniques," vol. 6, no. 2, pp. 103-111, 1993.

[15] L. G. Swan, V. I. J. R. Ugursal, and s. e. reviews, "Modeling of end-use energy consumption in the residential sector: A review of modeling techniques," vol. 13, no. 8, pp. 1819-1835, 2009.

[16] D. J. Livingstone, D. T. Manallack, and I. V. J. J. o. c.-a. m. d. Tetko, "Data modelling with neural networks: advantages and limitations," vol. 11, no. 2, pp. 135-142, 1997.

[17] F. Ziel, R. Steinert, and S. J. E. E. Husmann, "Efficient modeling and forecasting of electricity spot prices," vol. 47, pp. 98-111, 2015.

[18] P. S. database. (2019). Available: https://www.pecanstreet.org/dataport/

[19] F. Burden and D. Winkler, "Bayesian regularization of neural networks," in Artificial neural networks: Springer, 2008, pp. 23-42.

[20] AEMO. ACCESS MARKET PORTALS. Available: https://aemo.com.au/energy-systems/electricity/national-electricity-market-nem/data-nem/data-dashboard-nem

Retrieved from "https://projectswiki.eleceng.adelaide.edu.au/projects/index.php?title=Projects:2019s2-25202_Electricity_Consumers%27_Greenness:_A_Measure_to_Quantify_Electricity_Consumers%27_Sustainability_and_Efficiency&oldid=15014"