Abstract— Today, the adoption of new technologies has

increased power system dynamics significantly. Traditional

long-term planning studies that most utility companies perform

based on discrete power levels such as peak or average values

cannot reflect system dynamics and often fail to accurately

predict system reliability deficiencies. As a result, long-term

future continuous profiles such as the 8760 hourly profiles are

required to enable time-series based long-term planning studies.

However, unlike short-term profiles used for operation studies,

generating long-term continuous profiles that can reflect both

historical time-varying characteristics and future expected power

magnitude is very challenging. Current methods such as average

profiling have major drawbacks. To solve this challenge, this

paper proposes a completely novel approach to generate such

profiles for multiple generation types using Generative

Adversarial Networks (GAN). A multi-level profile synthesis

process is proposed to capture time-varying characteristics at

different time levels. Both Single-type GAN and a modified

Conditional GAN systems are developed. Unique profile

evaluation metrics are proposed. The proposed approach was

evaluated based on a public dataset and demonstrated great

performance and application value for generating long-term

continuous multi-type generation profiles.

Index Terms—Power System Planning, Generation Forecast,

Machine Learning

I. INTRODUCTION

raditonally, utility companies have commonly used

representative discrete values obtained from long-term

forecast such as peak, average or percentile values as input for

long-term system planning studies. Today, driven by new

technologies such as renewable energy, energy storage and

demand response, system dynamics has increased dramatically

[1]. As a result, system planning now requires much more

granular studies based on time-series data instead of discrete

values. For example, 8760 hourly profile is becoming a new

requirement for planning studies in today’s power industry

[1-6]. This level of data resolution can adequately model the

time-varying behaviours of loads and generations and the

combined effect to a power system. Based on such profiles,

accurate time-series based planning studies and simulations

M. Dong is with AESO in Canada; K.Xie and W.Li are with State Key

Laboratory of Power Transmission Equipment & System Security and New Technology, Chongqing University, No.174 of Sha Zhengjie Street,

Shapingba District, Chongqing, Zip Code: 400030 (e-mail:

mingdong@ieee.org;a kaiguixie@vip.163.com; wenyuanli630@gmail.com)

can be performed to realistically identify future system

deficiencies and derive proper planning decisions. However,

comparing to only a few values, creating 8760 generation

profiles is a much bigger challenge as the historical

time-varying characteristics are difficult to capture and

summarize as hard rules or assumptions. Currently, for

creating continuous generation profiles, the industry has made

the following attempts [7]:

Selecting one or making an average profile from a few

historical yearly profiles and applying it to future years.

This method is straightforward but has a few major

drawbacks: using one historical profile may not be

representative as other historical data is ignored; averaging

profiles can cause the loss of dynamic details that are

important for time-series analysis; using one historical or

the averaged profile can cause diversity issues. Profiles of

different years and/or generation sites will look the same

and are unrealistic. This method is difficult to incorporate

expected power magnitude that forecasting engineers can

predict based on their analysis of area load, new generation

capacity, production cost, renewable energy policy,

long-term weather senarios and etc. [6]

Randomly sampling daily profiles. This method randomly

picks a daily profile from an existing profile database and

scales it to match with the expected power magnitude. This

method may be applicable to some operation studies

when only separate daily profiles are required but is not

applicable to long-term continuous profiles. Thisis because

such multiple daily profiles cannot be direcly connected.

No data continuity will exist between two adjacent days

unless the starting and ending point values of each day are

manually altered for smooth transition, which is difficult to

do and also can damage the inherent profile patterns.

Furthermore, with limited existing daily profiles, this

method also lacks diversity, especially when profiles of

multiple future years are often required for long-term

planning studies.

Simulating generation behaviours for every future hour.

The method is difficult to apply because the simulation

models can be quite complex and costly to build and often

requires special commercial software or service[8]. The

accuracy relies on a large set of generations and external

assumptions that are often difficult to set. The results are

sensitive to assumptions and can vary significantly

between cases. In reality, the simulation is often unable to

Generating Long-term Continuous Multi-type

Generation Profiles using Generative

Adversarial Network Ming Dong, Senior Member, IEEE, Kaigui Xie, Senior Member, IEEE, Wenyuan Li, Life Fellow,

IEEE

T

reproduce historical data characteristics. Also, the

simulation process can be computationally expensive and

may need numerous trials to land on convincing results.

To address the above challenges, this paper proposes a

data-driven approach based on Generative Adversarial

Network (GAN) models to automatically capture the historical

time-varying characteristics of different types of generations

and more importantly, reproduce such characteristics when

generating future time-series profiles. In recent years, GAN

has become a hot topic in the AI research and application field.

It has been applied successfully to capture image

characteristics and create very similar images that human

cannot tell apart from the real ones. The idea of synthesizing

time-series generation profiles using GAN is inspired by such

accomplishments [9-11].

The current research on applying GAN to solve power

system problems is still at an early stage: [12] uses GAN to

generate missing PMU data for power system dynamic

security assessment; [13-15] uses GAN to generate building

and appliance loading data to augment limited training

datasets for relevant machine learning tasks; [16] uses GAN to

increase solar irradiance data to enhance weather classification

accuracy. As can be seen, [13-16] essentially tried to solve the

small sample learning problem by using GAN; [17] uses GAN

to extract characteristics from smart metering data and

replaces real smart metering data with synthesized smart

metering data so that the privacy of real smart metering data

can be protected; [18] uses GAN to add load forecast

uncertainty to the single-point deterministic forecast results

and hence make the load forecast result probabilistic.

Compared to the above research works, the research problem

we discuss in this paper is unique and is focused on producing

long-term continuous hourly profiles for multiple generation

types with the following specific requirements:

The generated profiles are able to reflect the historical

time-varying characteristics;

The generated profiles are able to reflect the expected

generation levels obtained from long-term generation

forecast;

The generated profile is for long-term use and in a

continuous form with hourly resolutions (i.e. 8760 points

a year);

The proposed method is applicable to different

generation types;

The proposed method is able to produce unlimited

number of different profiles to meet the diversity

requirement.

This paper is organized as follows: Section I explains the

research background, long-term continuous generation profile

requirements and presents literature review on relevant

research works; Section II proposes the multi-level profile

synthesis process which is a novel and critical process for

creating continuous long-term profiles by using GAN; Section

III proposes two GAN systems with different structures to

serve the purpose of generating profiles for multiple

generation types; Section IV focuses on the explanation of the

training process for GAN models and proposes a new loss

function for Multi-type GAN; Section V further proposes

evaluation metrics and provides detailed validation of the

proposed method on a public long-term generation dataset

with comparison to the traditional average profiling method. It

also briefly discusses a Monte-Carlo based post-processing

step in order to add random generation outages to the profiles.

The evaluation demonstrates great performance and

application value of the proposed method. In Section VI, some

future improvement opportunities are also identified and

discussed

II. MULTI-LEVEL PROFILE SYNTHESIS PROCESS

In order to produce long-term continuous generation profile,

this section proposes a unique multi-level process that can use

different GAN models to capture and reproduce time-varying

characteristics at different time levels.

A. Multi-level Synthesis Process

For long-term profile such as the 8760 hourly profile, it is

important to recognize that within the whole-year’s time-series

data, different time-varying characteristics can reveal at

different time levels. This paper proposes a unique and

powerful multi-level profile generation process to capture the

time-varying characteristics at different time levels. We

illustrate this process based on one-year time but multiple

years can be produced in the same way and joined together.

As shown in Figure 1, at the first level, a yearly profile in

which every data point represents a weekly generation level is

generated (52 points in total). The weekly generation level can

be average weekly generation, for example; then for every

week, a weekly profile is generated with respect to the weekly

generation levels that the yearly GAN model have generated at

the first level. The weekly profile can be generated directly

(168 points) or on a daily basis (7 points). If produced on a

daily basis, a third profile level is required for generating daily

profiles which have 24 points for 24 hours and align with the

daily generation levels that the second-level weekly GAN

model has generated.

Fig.1. Illustration of two-level and three-level profile synthesis

Unlike the yearly profiles, weekly and daily profiles have

many more time-varying characteristics and therefore should

be learned with dedicated GAN models. The proposed

multi-level profile synthesis process is a powerful idea as it

means the yearly GAN model should only focus on the big

picture, i.e. how generation level varies from week to week

throughout a year; then the weekly and daily models can focus

on the more short-term time-varying characteristics within a

week or a day. In the end, time-varying characteristics of the

whole spectrum can be captured and reproduced by GAN

models at different time levels.

B. The “Starting-point” Trick

The multi-level synthesis process allows the use of multiple

GAN models to capture time-varying characteristics at

different time levels. The weekly and daily GAN models will

generate weekly and daily profiles one after another.

However, if the models generate every period separately with

no coordination, it will be impossible to ensure the data

continuity between every two adjacent periods. For example,

the last point of weekly profile is 50MW while the first

point of weekly profile is 70MW – directly linking

these two weeks together will yield a 20MW jump that is

unrealistic. To address such an issue and seamlessly connect

two adjacent periods, we propose and introduce the

“starting-point” trick.

The “starting-point” trick means when GAN is used to

generate a profile, the ending point of the last period is also

included as the starting point of the current GAN generated

period. Instead of generating points (the length of the

period), the GAN model now generates points. The

profiles of the last period and current period to be generated

have an overlapping point so that the two adjacent periods can

be connected seamlessly.

As shown in Fig.1, the “starting-point” can also help to

reflect the expected magnitude. The profiles used for training

a GAN are normalized generation profiles. Therefore they are

essentially shape profiles without power information directly

incorporated. However, when a certain type of period

magnitude such as the average generation power is expected,

the ratio between the starting point and the period magnitude

can be calculated. We define this ratio as the starting-point

ratio . For example, when using the average generation

power, is defined as:

∑

where is the magnitude value of the starting point; to

is the new period that actually needs to be generated.

Since the starting point is also the ending point of the last

period and the power magnitude of the ending point is already

known after generating the last period, the expected period

magnitude can be incorporated as long as the generated profile

aligns with the expected

Fig.2. Illustration of the “starting-point” trick

When generating a new period profile, the expected is

firstly calculated according to the ending point of last period

and the current expected period magnitude. Then there are two

methods to generate a profile that aligns with the expected the GAN model can generate profiles continuously until a

generated profile gets a with an error less than a certain

threshold; alternatively, the GAN model can generate a large

amount of profiles such as 10,000 profiles upfront before the

entire profile generation task starts. This will form up a

baseline profile database. Then the profile with the closest

can be selected from this database. It should be noted that

these two methods can also be combined for use. Finally, once

a profile with the expected is generated, it can be easily

converted to power magnitude based profile given the known

power magnitude of the starting point.

C. The “Duty-cycle” Trick

It should be noted the starting point trick can apply to most

generation types which have continuous output. However,

there is also a “duty-cycle” type of generation such as some

simple-cycle generations which are studied and illustrated in

Section V of this paper. In such cases, the generation can be

shut down frequently and the minimum generation output can

drop to zero. The continuity between two adjacent periods is

not required because there are gaps by its nature. In such

cases, an expected power magnitude can be converted to an

expected duty-cycle factor. Then summing up every point in

the target period in the generated profile (0-1 normalized) will

get the duty-cycle factor. Finally, similar to the

“starting-point” trick, the algorithm only needs to find out a

GAN generated profile that aligns with the expected

duty-cycle factor in order to reflect the expected magnitude.

Mathematically, the duty cycle factor is calculated as

below:

∑

Where to are the magnitude values of points in

the period; is the generation capacity.

III. TWO GAN SYSTEMS

This section explores two different GAN systems that can

deal with the proposed problem: Single-type GAN and

Multi-type GAN.

A. Single-type GAN

A Single-type GAN can only produce profiles for one type

of generation such as wind generation or coal generation. It

also needs to be trained based on one type of generation’s

historical real data. Its structure is shown in Fig.3. As can be

seen, it has a generator and a discriminator:

A generator is a neural network with deconvolutional

layers that can generate time-series profiles. It has a

multi-input and multi-output (MIMO) structure which

takes in a latent vector as input and produces a time-series

vector as the output. The latent vector is just a vector

consisting of noise elements that are typically sampled

from normal distributions.

A discriminator is a convolutional neural network that

serves as a binary classifier to judge if an input time-series

vector is a real or generated profile. To train the

discriminator, a combination of real profiles and generated

profiles should be fed into the discriminator with existing

known “real” or “fake” labels.

The whole GAN training process works as below [19]: at

the beginning the generator has no clue on how to generate a

realistic time-series profile so it generates a time-series profile

almost randomly; the discriminator is trained with labeled real

and generated profiles. Since the ultimate purpose of the

generator is to generate realistic profiles, the generator is

expected to eventually confuse the discriminator and make the

discriminator misclassify the generated profiles as real

profiles. However, because at the early stage the generator

generates almost randomly, the discriminator will produce a

considerable network loss to train the generator. After the

generator gets trained after one iteration, the generator will get

slightly better at generating time-series. At the same time, the

discriminator will also be trained with the newly generated

profiles and real profiles, trying to still differentiate them. A

new loss from the discriminator will be generated and

continue to guide the generator to improve. This process

continues iteratively and both the generator and discriminator

are getting better. The iteration process can stop after no

significant improvement is observed on the generator output

or when the discriminator stops to effectively differentiate a

real and a generated profile. In the end, the generator model is

kept as the final usable part of the GAN model for generating

future profiles.

Fig.3. Structure of a Single-type GAN system

B. . Multi-type GAN

Since the planning area often has multiple types of

generations, to use Single-type GAN to generate different

types of generation profiles, multiple GAN models have to be

developed, trained and used separately. For more convenient

implementation, this paper also explores the possibility of

using one GAN model to generate different types of

generation profiles. In the GAN research field, this is called a

Conditional GAN problem [20]. A classic Conditional GAN

structure is shown in Fig.4. Compared to Single-type GAN

model, a fundamental structure difference is that an additional

condition input is introduced and combined with the latent

vector and fed into both the generator and discriminator. The

discriminator now not only needs to judge if the generated

profile is true or fake but also needs to judge if it matches with

the expected condition requirement. Conditional GAN has

been widely used for context based image generation,

image-to-image translation and etc. in recent years.

Fig.4. Structure of a classic Conditional GAN system

Similar to Single-type GAN, the classic Conditional GAN

has only one output value indicating real or fake. 1 indicates

real and 0 indicates fake. However, the requirements of being

real or fake in Conditional GAN are different from Single-type

GAN. The output value should be 1 only when the generated

profile looks real and at the same time matches with the

expected condition; if the generated profile looks fake, the

value should be 0; if the generated image looks real but does

not match with expected condition, the value should also be 0.

As can be seen, this output value 0 is ambiguous now because

it includes two meanings. As a result, the discriminator can

become very strict - even if the generator somehow is able to

generate a close-to-real profile, it can still get significant

punishment simply because the generated profile does not

match with the condition. This could cause the training to get

stuck at the early stage when generator is weak and also lead

to the famous mode collapse problem for GAN. When mode

collapse happens, the generated profiles look completely or

partially similar. We tested classic Conditional GAN for the

discussed application and ran into mode collapse very

frequently. An example is shown in Fig. 5. To solve this

problem, researchers have tried establishing a third neural

network in parallel with the discriminator neural network to

produce an additional loss that only focuses on describing the

match of conditions. Then the two losses need to be properly

combined as one for the discriminator [21-22]. However, this

approach significantly increases the implementation and

training complexity of the entire GAN system.

Fig.5. A mode collapse example of generated wind generation profiles

Instead, this paper proposes another modified structure to

reduce the ambiguity of the discriminator output – the key

philosophy is to diversify the discriminator’s judgement and

remove output ambiguity. A modified Condition GAN named

Multi-type GAN is proposed and tested. Its structure is shown

in Fig. 4. There are three main changes compared to the

classic Conditional GAN:

Fig.5. Structure of the proposed Multi-type GAN system

The discriminator is changed from a binary classifier to a

multi-class classifier which allows the output to be a

vector instead of a scalar value. To indicate the

generation type, we adopt the easy-to-implement one-hot

encoding. For example, assuming a planning area

contains three types of generations: wind, coal and gas

generations. First, a three-element Type Code Vector

can be adopted and [1,0,0], [0,1,0] and [0,0,1] can

represent the wind, coal and gas generations respectively.

v instructs the generator to generate a profile according

to the desired generation type; second, a four-element

Type Code Vector v’ that indicate the generated profile

looks like real wind, real coal, real gas generation or does

look like any of the three above is adopted.

A softmax layer is added to the multi-class classier to

describe the probabilities of each category [ ]. This aligns

with the one-hot encoding input very well as the

summation of all categorical probabilities is always equal

to 1. The corresponding loss function is explained in

detail in Section IV-C.

The double feed of condition information to the

discriminator is no longer needed. This also reduces the

system complexity.

IV. TRAINING OF THE GAN MODELS

This section discusses training issues for the two GAN

systems.

A. Normalized Real Profiles

Since the GAN generated profile only focuses on the profile

shape, similar to load shape profile analysis in [23], a real

generation profile should be normalized to the range of 0-1 by

using the following equation:

where is any data point on the raw time-series; for

generations with continuous output, is the peak value of

the profile; for “duty-cycle” type of generations, is the

generation capacity.

B. Minimax and Wasserstein GAN Losses

[19] introduces the first widely used GAN loss function:

[ ] [ ( )

where E is the expected value of the term in its bracket; is a

latent vector and is the generator's output when given ;

is a real historical profile; is the discriminator’s

output reflecting how probable the given profile is real.

Ideally, it should be 1 for a real profile and be 0 for a

generated profile. This loss is also called Minimax loss

because the generator tries to minimize it while the

discriminator tries to maximize it

It was found later that the Minimax loss function can cause

GAN to get stuck at the early training stage. Therefore, [24]

proposes another widely used loss function called Wasserstein

GAN loss (WGAN). The discriminator’s loss function is:

( ) The generator’s loss function is:

( )

In WGAN, both generator and discriminator try to

maximize their loss functions. WGAN is used for the

Single-type GAN system.

C. Proposed Multi-type GAN Loss

As discussed in Section III, a Multi-type GAN has a

multi-class classifier. The proposed loss functions are

carefully designed to mimic the binary Wasserstein GAN loss

and still reflect the multi-class classification features. The loss

of the discriminator is:

[ ] [∑ ( )

]

where is the generation type index; is the number of

generation type. For real profiles, it is expected that their

corresponding probability values in the discriminator Type

Code vector ’ are as high as possible; for generated profiles,

it is expected that the average of all real-type probability

values (1st to values) in the discriminator Type Code

vector is as low as possible.

The loss function of the generator is:

[ ( )] (8)

Both discriminator and generator are trying to maximize

their loss functions.

V. VALIDATION AND EVALUATION

We tested the above proposed methods on a publicly

available dataset [25]. The dataset contains historical 10-year

hourly generation data for multiple coal, simple-cycle gas and

wind generation units from a planning area in Canada. This

section presents and discusses the results by using the

proposed methods. A few metrics are also proposed to

quantitatively evaluate and compare the quality of the

generated profiles.

A. Visual Analysis

Coal, simple-cycle gas and wind generation profiles are

generated by both single-type GAN and Multi-type GAN

models. Since weekly profiles contain abundant time-varying

characteristics and a long duration, we choose weekly profiles

to illustrate the differences between the real and GAN

generated profiles. For comparison purpose, 5 randomly

selected profiles are presented for each category. As can be

seen from Fig.6, some observations can be made:

Three generation types have three distinct output

patterns. Coal generation generates more steadily with

fluctuation around a relatively narrow range;

simple-cycle gas generation is a typical duty-cycle based

generation; wind generation fluctuates more widely and

frequently by nature. Such patterns have been

successfully captured by both Single-type GAN and

Multi-type GAN models.

The output of both Single-type GAN and Multi-type

models resemble the real profiles. However, compared

with single-type GAN, the quality of Multi-type GAN

seems to be lower with sometimes high-frequency noises

at some places of the profiles.

Although visual comparison can provide a straightforward

validation on the results, this section further proposes a few

Fig.6. Visual Comparison of 5 randomly selected weekly profiles for each category

insightful metrics to quantify and compare the profile quality

from different perspectives.

B. Quality Metric on Time-varying Characteristics

To evaluate whether the generated profiles have

successfully captured the original time-varying characteristics,

Kullback–Leibler Divergence (KL-divergence) has been

adopted. KL divergence is a metric to describe how one

probability distribution is different from another.

KL-divergence has often been used in GAN applications to

evaluate the similarity of generated data and real data [17].

Mathematically, KL-divergence can be calculated as below:

∫

where is the probability distribution of a real profile;

is the probability distribution of a generated profile.

is greater than 0. When is zero, it

means the two probability distributions are exactly the same.

Therefore, a smaller indicates that the generated

profile has better captured the time-varying characteristics of

the real profile.

C. Quality Metric on Power Magnitude Error

To evaluate whether the generated profile has successfully

captured the expected power magnitude for a year, absolute

percent error (APE) can be adopted. It is given as:

| ̂

|

where is the expected power magnitude; ̂ is the generated

power magnitude; Typically, and ̂ can be peak, average

or percentile values depending on the generation forecasting

preference.

D. Quality Metric on Profile Diversity

As a requirement listed in Section I, the GAN models

should be able to generate profiles with different shapes. In

other words, the diversity on the generated profiles should be

high. We propose the following metric based on the concept of

standard deviation to quantify the diversity of daily profiles

normalized to between 0 and 1.

∑√

∑ ̅

where is the number of data points in a generated profile;

is the number of generated profiles; is the value of the

data point at the generated profile; ̅ is the average

value of the data point across all generated profiles.

E. Evaluation Summary

We applied the above proposed quality metrics to both

single-type and Multi-type GAN models for both the two-level

and three-level processes for the three generation types in the

dataset. To reveal the advantages of the GAN methods, the

traditional average profiling method mentioned in Section I is

compared. The method of randomly sampling daily profiles is

not considered because it is unable to establish realistic

continuous profiles due to the data continuity issue between

adjacent periods. To start generating an 8760 profile, a starting

power and an expected average generation power need to be

assumed. After generating the profiles, for each generation

type, real 8760 profiles are compared to the generated profiles

and their KL-divergence is calculated. For reference purpose,

the KL-divergence of two groups of real profiles is also

calculated and taken; is calculated against the expected

average generation power. Table I shows the summary of the

KL-divergence and average values for 10 generated

profiles under each method. For S, daily profiles are used for

comparison because in reality each day is supposed to look

different. is applied to 2000 normalized daily profiles

generated from the weekly (two-level) or daily (three-level)

GAN models. of 2000 real normalized daily profiles is

calculated as reference. We also included the total training

time that is required to establish the models in comparison.

The tests were done on a computer with Intel i7-6700K 4.0

GHz CPU with 16.0GB RAM memory. The software

environment is Python 3.7 with Tensorflow 2.0.

TABLE I: SUMMARY OF PROFILE QUALITY EVALUATION

Method Generation KL-

divergence

APE Training

Time(s)

Single-type

GAN (Three-level)

Coal 0.62 0.32 73.5 181

Simple-cycle Gas

0.09 1.69

14.2 646

Wind 0.02 0.12 20.1 450

Multi-type

GAN (Three-level)

Coal 0.71 0.75 57.6

661

Simple-cycle

Gas 0.14

1.10

14.5

Wind 0.03 1.89 18.9

Single-type

GAN

(Two-level)

Coal 0.69 2.04 67.2 144

Simple-cycle

Gas 0.13

0.10

14.4 495

Wind 0.02 1.05 19.7 276

Multi-type

GAN

(Two-level)

Coal 0.78 1.88 53.3

673 Simple-cycle

Gas 0.17

0.95

14.3

Wind 0.04

2.13 18.8

Averaged

profiles

Coal 1.03 5.33 41.8

N/A Simple-cycle

Gas 0.3

7.56 11.1

Wind 0.07 6.30 17.3

Real Profiles

Coal 0.52 N/A 81.2

N/A Simple-cycle

Gas 0.07

N/A 16.5

Wind 0.01 N/A 21.1

A few observations can be made from Table I:

Overall, the proposed GAN systems can reflect

time-varying characteristics, expected forecast magnitude

and keep good profile diversity at the same time. This

can be told by comparing to the results of real profiles

and averaged profiles.

Overall, two-level process has slightly higher

KL-divergence than the three-level process (with only

one result exception). However, it has a better .

Also, it is simpler to implement because it can eliminate

the need to train the third-level daily GAN models.

Single-Type GAN has better KL-divergence, and

values than Multi-type GAN. This is because a

single-type GAN model is trained separately and there is

no interference from other generation types during

training. However, the drawback is that more models

have to be developed, trained and stored for application.

The effort can be significant with more generation types

in the planning region. As the table shows, just the total

training time itself is higher than Multi-type GAN.

The averaged profile has a much higher KL-divergence

compared to real and GAN generated profiles. This is

because the averaged profile eliminates time-varying

details and distorted the inherent patterns. It also has a

significantly lower than the real and GAN generated

profiles. This is because the averaged profile lacks

diversity and there are many repeating daily profiles out

of the total 2000 tested daily profiles.

F. Monte-Carlo Post-processing

The profiles generated by the above GAN models are

continuous. In reality, sometimes generations may take

outages due to maintenance or other reasons. If this needs to

be reflected in the profiles, one can simply perform a

sequential Monte-Carlo simulation (MCS) over the forecasting

window. The sequential MCS can incorporate certain expected

outage rate and outage duration probability distribution. Fig.7

shows an example of coal generation profile with a 2/year

outage rate and an average of 1-week outage duration during a

year.

Fig.7. A final 8760 wind generation profile with 2 outages in a year

VI. CONCLUSIONS AND DISCUSSIONS

This paper proposed a completely novel approach for

generating long-term continuous multi-type generation

profiles using GAN models. The paper firstly analyzes the

drawbacks of traditional methods and summarizes the

requirements. The novelties of the paper can be summarized

as below:

it proposes a unique multi-level profile synthesis process

with the “starting-point” trick and “duty-cycle” trick that

is able to capture time-varying characteristics at different

time levels and reflect the expected power magnitude;

it proposes a new Conditional GAN structure called

Multi-type GAN and its unique loss function

it proposes a series of metrics for evaluating the quality

of the GAN generated profiles.

The approach was applied to a public dataset. Different

variations were implemented and compared from different

requirement perspectives. The result shows that the proposed

approach is able to:

generate realistic profiles that capture time-varying

characteristics at different time levels for different types

of generations;

generate profiles with specific magnitude that complies

with long-term forecasting expectation;

and generate diverse profiles that are never repeated.

The proposed approach is an effective solution for

generating long-term continuous generation profiles for

multiple types of generations in order to serve the increasing

need of dynamic long-term system assessment and

simulations. Based on the existing effort and outcome, there is

still an improvement room in the future:

More advanced Conditional GAN systems can be

investigated such as using a third neural network for

condition match judgment and using special embedding

layers to incorporate condition information [26];

Generate profiles with more granular condition

information such as specific operation characteristics in

addition to general type. For this purpose, a Style-GAN

based solution may be needed for further development

[27].

REFERENCES

[1] J. Kays, C. Rehtanz, “Planning process for distribution grids based on flexibly generated time series considering RES, DSM and storages”, IET Generation, Transmission & Distribution, vol.10, pp 3405–3412, 2016.

[2] J. Kays, A. Seack, U. Häger, “The potential of using generated time series in the distribution grid planning process,” Proceedings of the CIRED 23rd International Conference on Electricity Distribution, Lyon, France, June 15–18, 2015.

[3] B.A. Frew, W.J. Cole, Y. Sun, T.T. Mai, and J. Richards, “8760-based method for representing variable generation capacity value in capacity expansion models (No. NREL/CP-6A20-68869),” National Renewable Energy Lab, Golden, CO, United States, 2017.

[4] F. Schäfer, J.H. Menke, F. Marten and M. Braun, “Time Series Based Power System Planning Including Storage Systems and Curtailment Strategies”, Proceedings of the CIRED 2019 25th International Conference on Electricity Distribution, Madrid, Spain, June 3-6, 2019

[5] Independent Electricity System Operator, “Preliminary 2019 Long Term Demand Forecast”,[Online].Available: http://ieso.ca/-/media/Files/IESO/Document-Library/engage/20-year-planning-outlook/Preliminary-2019-Long-Term-Demand-Forecast_Stakeholder-Engagement_UpdatedFeb1.pdf?la=en

[6] Alberat Electric System Operator, “AESO 2020 Long-term Transmission Plan”, [Online].Available: https://www.aeso.ca/assets/downloads/AESO-2020-Long-termTransmissionPlan-Final.pdf

[7] F. Elkarmi, ed., Power System Planning Technologies and Applications: Concepts, Solutions and Management, IGI Global, 2012.

[8] Energy Exemplar, “Aurora Forecasting Software”, [Online].Available:https://energyexemplar.com/solutions/aurora/

[9] A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, and A.A. Bharath, “Generative adversarial networks: An overview,” IEEE Signal Processing Magazine, 35(1), pp.53-65, 2018.

[10] K. Wang, C. Gou, Y. Duan, Y. Lin, X. Zheng and F.Y. Wang, “Generative adversarial networks: introduction and outlook”, IEEE/CAA Journal of Automatica Sinica, vol.4, no.4, pp.588-598, 2017.

[11] M. Zamorski, A. Zdobylak, M. Zięba, and J. Świątek, “Generative Adversarial Networks: recent developments,” International Conference on Artificial Intelligence and Soft Computing, Springer, Cham, 2019, pp. 248-258,

[12] C. Ren and Y. Xu, “A Fully Data-Driven Method Based on Generative Adversarial Networks for Power System Dynamic Security Assessment With Missing Data,” IEEE Transactions on Power Systems, vol. 34, no. 6, pp. 5044-5052, Nov. 2019

[13] S. E. Kababji and P. Srikantha, “A Data-Driven Approach for Generating Synthetic Load Patterns and Usage Habits,” IEEE Transactions on Smart Grid, vol. 11, no. 6, pp. 4984-4995, Nov. 2020.

[14] C. Tian, C. Li, G. Zhang, and Y. Lv, “Data driven parallel prediction of building energy consumption using generative adversarial nets,” Energy and Buildings, vol.186, pp.230-243. 2019.

[15] M.N. Fekri, A.M. Ghosh, and K. Grolinger, “Generating energy data for machine learning with recurrent generative adversarial networks,” Energies, vol.13, no.1, p.130, 2020.

[16] F. Wang, et al., “Generative adversarial networks and convolutional neural networks based weather classification model for day ahead short-term photovoltaic power forecasting,” Energy Conversion and Management, vol.181, pp.443-462, 2019.

[17] Z. Wang and T. Hong, “Generating realistic building electrical load profiles through the Generative Adversarial Network (GAN),” Energy and Buildings, vol. 224, p.110299, 2020.

[18] Y. Wang, G. Hug, Z. Liu and N. Zhang, “Modeling load forecast uncertainty using generative adversarial networks,” Electric Power Systems Research, vol.189, p.106732, 2020.

[19] Goodfellow et al., "Generative adversarial nets," Advances in Neural Information Processing Systems, pp. 2672-2680, 2014.

[20] M. Mirza and S. Osindero, “Conditional generative adversarial nets,”arXiv preprint:1411.1784,2014.

[21] A. Odena, C. Olah and J. Shlens, “Conditional image synthesis with auxiliary classifier GANs,” Proceedings of the 34th International Conference on Machine Learning,Sydney, Australia, 2017,pp. 2642-2651.

[22] T. Miyato, T. Kataoka, M. Koyama and Y. Yoshida, “Spectral normalization for generative adversarial networks,”arXiv preprint:1802.05957, 2018.

[23] M. Dong, “A Data-driven Long-term Dynamic Rating Estimating Method for Power Transformers,” IEEE Transactions on Power Delivery, Early Access

[24] M. Arjovsky, S. Chintala and L. Bottou, “Wasserstein GAN,” arXiv preprint:1701.07875, 2017.

[25] Utility Analtyics Networks, “Dataset: Area 10-Year Generation Data in Canada”, [Online].Available: https://www.kaggle.com/utilityanalytics/10year-generation-data

[26] Y.Ren, Z.Zhu, Y.Li and J.Lo, “Mask Embedding in conditional GAN for Guided Synthesis of High Resolution Images”, arXiv preprint:1907.01710, 2019.

[27] T. Karras, S. Karras and T. Aila, “A style-based generator architecture for generative adversarial networks,” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4401-4410, 2019.

Ming Dong (M’13-SM’18) has been working a senior research engineer and data scientist in Canadian power industry for 8 years. He received his Ph.D degree from

University of Alberta and the Certificate of Data Science and Big Data Analytics

from Massachusetts Institute of Technology. He is an editor of IEEE

TRANSACTIONS ON POWER DELIVERY and CSEE JPES. His research focus

is the applications of AI and big data analytics to power system planning and

operations.

Kaigui Xie (M'10–SM'13) received the Ph.D. degree in power system and its

automation at Chongqing University in 2001. Currently, he is a full professor in the

School of Electrical Engineering, Chongqing University. His main research

interests focus on the areas of power system reliability, planning and analysis. He is

the author and coauthor of over 200 academic papers and five books. Prof. Xie was

an editor of the IEEE TRANSACTIONS ON POWER SYSTEMS and a member of

editorial board of Electric Power Components and Systems.

Wenyuan Li (F’02-LF’18) is currently a professor with Chongqing University, Chongqing, China. His research interests include power system

asset management, planning, operation, optimization, and reliability

assessment. He is a Fellow of the Canadian Academy of Engineering and the Engineering Institute of Canada, and a Foreign Member of the Chinese

Academy of Engineering. He was a recipient of several IEEE PES awards, including the IEEE PES Roy Billinton Power System Reliability Award, in

2011, and the IEEE Canada Electric Power Medal, in 2014.