Popular in Mathematics

Formal Sciences

Diameter rigidity for Kähler manifolds with positive bisectional curvature

Indian Institute of Science Ved Datar

Formal Sciences

Rank-Based Multivariate Sarmanov for Modeling Dependence

McMaster University Anas Abdallah

Formal Sciences

The Hybrid High-Order Method for Polytopal Meshes

Monash University

Formal Sciences

Financial Informatics: An Information-Based Approach to Asset Pricing

University of Surrey

Formal Sciences

Multiplicative orders of Gauss periods and the arithmetic of real quadratic fields

University of Newcastle

Formal Sciences

Some Numerical Methods for the Hele-Shaw Equations

University of Massachusetts at Amherst

Formal Sciences

Geometric stabilisation via p-adic integration

University of Toronto

Probabilistic Forecasting of Wind and Solar Farm Output

View more insights in Sciences - Mathematics - Applied Mathematics - Formal Sciences

View more figures from University of South Australia

View the page of John Boland

Read the Study

Next in Applied Mathematics

Microswimmer Propulsion by Two Steadily Rotating Helical Flagella

| University of Waterloo

John Boland discusses statistical forecasting tools to generate forecasts with prediction intervals, trailing them on one wind and one solar farm.

Image courtesy of interviewee. September 20, 2024

Log-in or Sign-up to Faculti
Currently viewing this subject insight as a guest. You have insight(s) remaining for this month. Login to view 8000+ figures on the platform.

Generating Workspace...

Export as PDF

This video video discusses the importance of probabilistic forecasting in energy generation from renewable sources like solar and wind farms, particularly in the context of the Australian energy market. It emphasizes that understanding the error bounds around energy forecasts is crucial for decision-making by both producers and grid operators. Narrow error bands signify confidence in predictions, allowing for more reliable integration into the grid, while wider bands may lead to the risk of penalties for misforecasting.

The speaker shares a scenario in which a solar or wind farm must balance the output to the grid with the possibility of underestimating or overestimating generation, which could incur costs. Therefore, accurate and sharp probabilistic forecasts can enhance the financial decisions made by generators.

This video video also highlights recent experiments using simple statistical models, which have proven effective for short-term forecasting, compared to more complex machine learning methods. These forecasts help evaluate the value and reliability of the predicted energy output, supporting operational flexibility and profitability.

Additionally, it mentions a research project funded by the Australian Renewable Energy Agency, focusing on improving five-minute forecasting for solar and wind outputs, and outlines a method for transforming forecasting errors into a more useful form, demonstrating a novel approach that outperforms traditional quantile regression.

Finally, the text concludes with thoughts on future developments in market mechanisms for renewables, suggesting that established prediction methodologies can facilitate financial decisions and operational efficiencies in an increasingly renewable-driven electricity market.

Language:

Export as PDF

Export as PDF
Here are five key points from the video:

1. **Importance of Probabilistic Forecasting**: The video emphasizes the significance of knowing the bounds around energy generation forecasts (e.g., from solar and wind farms) in the convideo of the Australian energy market. Narrower bounds around forecasts lead to more confident decisions by both energy generators and grid operators, while wider bands may prompt generators to store energy in batteries to avoid penalties for misforecasting.

2. **Consequences of Misforecasting**: Incorrect forecasts can lead to financial penalties due to the "cause or pay" system in the Australian energy market. Under-forecasting can force grid operators to source additional power, while over-forecasting may disrupt other generators' supply, underlining the necessity for accurate and reliable forecasting methods.

3. **Efficiency of Simple Forecasting Tools**: Recent experiments showed that simplified statistical models can effectively match or even outperform complex machine learning tools for short-term forecasting (e.g., five-minute forecasts), particularly in environments where quick decision-making is crucial, such as in electric energy markets.

4. **Value of Forecasting Beyond Accuracy**: The exploration of forecasting focuses not just on the skill or accuracy of the forecasts but also on their practical utility. Work is being done to assess the value of forecasts in operational and financial terms, especially on Reunion Island as well as in the convideo of Australia's national electricity market, highlighting the importance of critically evaluating how useful forecasts are in real-world applications.

5. **Innovative Forecasting Techniques**: The video discusses an innovative approach using transformations of errors with ARCH and GARCH methods to forecast the variance of forecasts. This methodology suggests the possibility of achieving sharper prediction intervals, which is essential for effective energy market operations. The concept of quantile regression is also mentioned, detailing how it can be used to forecast different percentiles to provide a comprehensive view of potential forecasting errors.

Export as PDF
Here are key quotes extracted from the video you :

1. **Forecasting Importance**:
- "It's my belief that instead of just forecasting the expected... it's better if you know the bounds around that forecast."

2. **Confidence in Forecasting**:
- "If the bands are narrow... then you can be sort of more confident that you can send that off to the grid."

3. **Implications of Misforecasting**:
- "If your forecast is a hundred, but you produce 80 or 120, it may mean that the grid operator will actually penalize you for misforecasting."

4. **Energy Market Mechanics**:
- "There's a thing called cause or pays in the... energy market in Australia."

5. **Value of Probabilistic Forecasts**:
- "If you have probabilistic forecasts... you might have to be flexible in the operation of the grid."

6. **Simplicity in Forecasting Tools**:
- "...sometimes people... believe that they might have to use very powerful tools... they don't necessarily check to see if simple tools will work as well."

7. **Experimentation with Forecasting**:
- "...we've just run a recent experiment... using very simple statistical models for a short time span."

8. **Forecasting Skill vs. Usefulness**:
- "It's not just good enough to think about the skill of the forecasting method, but... how useful it is, is it actually of use."

9. **Impact of Innovation**:
- "The nice thing is that... recent papers talk about... how useful it is in the sense that we've just explained for the generator and the operator of the system as well."

10. **Future of Renewable Forecasting**:
- "By 2027, a hundred percent of electricity will come from renewables... over the year, the total generated will be a hundred percent of the supply."

11. **Probabilistic Forecasting Benefits**:
- "Putting air bounds on the forecast will help them make their decisions on the financial side of things better."

12. **Combining Approaches**:
- "One place... find that they were very useful was... blending those forecasts... trying to figure out what... different types of models might work better."

These quotes encapsulate the major themes regarding forecasting in the convideo of renewable energy generation, emphasizing the importance of accurate predictions, the balance between simple and complex tools, and the operational implications of forecasting for energy producers and grid operators.

X-TIMESTAMP-MAP=LOCAL:00:00:00.000,MPEGTS:90000

1
00:00:00.075 --> 00:00:02.485
It's my belief that instead of just

2
00:00:03.405 --> 00:00:07.405
forecasting the expected, generation, shall we say,

3
00:00:07.405 --> 00:00:10.245
from a solar, wind farm in the next five minutes,

4
00:00:10.335 --> 00:00:11.885
which is the timeframe

5
00:00:11.885 --> 00:00:14.245
that the energy market works on in Australia,

6
00:00:14.625 --> 00:00:19.525
or even, you know, a longer period, it's better if you know

7
00:00:20.425 --> 00:00:22.445
the bounds around that forecast.

8
00:00:22.645 --> 00:00:24.565
I think are they, are they narrow?

9
00:00:25.105 --> 00:00:28.885
Is your forecast really sharp or are they wide?

10
00:00:29.185 --> 00:00:33.285
Is your forecast, you know, a bit loose, shall we say?

11
00:00:33.665 --> 00:00:37.765
Uh, because that can help the decision making both on,

12
00:00:38.185 --> 00:00:42.365
the generator's side and on the operator of the system

13
00:00:42.425 --> 00:00:43.925
or the grid, for instance.

14
00:00:44.085 --> 00:00:47.965
I think with the, on the generator side, let's,

15
00:00:48.065 --> 00:00:49.245
have a little scenario.

16
00:00:49.385 --> 00:00:51.645
For instance, say you've got a, a solar farm

17
00:00:51.745 --> 00:00:55.445
and you have a battery, attached to it as well,

18
00:00:55.655 --> 00:00:57.845
which is becoming more and more common,

19
00:00:57.905 --> 00:00:59.125
or a wind farm as well.

20
00:00:59.955 --> 00:01:04.005
Okay? So you predict in the next five minutes,

21
00:01:04.505 --> 00:01:07.645
you're going to output a hundred megawatts

22
00:01:07.645 --> 00:01:08.765
from the, from the farm.

23
00:01:09.115 --> 00:01:12.265
Okay? Now, if the

24
00:01:13.935 --> 00:01:16.045
bands are, the bands around,

25
00:01:16.065 --> 00:01:20.405
or the, the possible error around that is small, okay,

26
00:01:20.835 --> 00:01:22.805
then you can be sort of more confident

27
00:01:22.805 --> 00:01:24.565
that you can send that off to the grid.

28
00:01:25.075 --> 00:01:28.765
Because if you mis forecast, in other words,

29
00:01:28.825 --> 00:01:31.525
if your forecast is a hundred, but you produce 80

30
00:01:32.025 --> 00:01:34.605
or 120, it may mean

31
00:01:34.605 --> 00:01:37.925
that the grid operator will actually penalize you

32
00:01:37.945 --> 00:01:39.325
for mis forecasting.

33
00:01:39.515 --> 00:01:40.845
There's a thing called cause

34
00:01:40.845 --> 00:01:44.165
or pays in the, energy market in Australia,

35
00:01:44.505 --> 00:01:46.925
for instance, if you under forecast, they have

36
00:01:46.925 --> 00:01:49.725
to bring more power on the grid from somewhere else.

37
00:01:50.425 --> 00:01:52.125
You basically have to pay for that fact.

38
00:01:52.945 --> 00:01:57.125
And if you, forecast too low, right?

39
00:01:57.635 --> 00:02:00.325
They may have to cut somebody else off, okay?

40
00:02:00.325 --> 00:02:04.365
That sort of thing. And so if your bands are wide

41
00:02:04.365 --> 00:02:08.365
around your expected output, you may decide,

42
00:02:08.585 --> 00:02:12.845
I'd rather put that, electricity into my battery now

43
00:02:13.545 --> 00:02:16.165
and then send it to the grid when I know exactly

44
00:02:16.275 --> 00:02:20.245
what I can get out from the battery rather than put it into

45
00:02:20.245 --> 00:02:22.365
the grid and risk a penalty.

46
00:02:22.835 --> 00:02:26.165
Okay? But if they're narrow, the bands around the forecast,

47
00:02:26.345 --> 00:02:30.925
you're confident it's okay on the operator side of the,

48
00:02:30.945 --> 00:02:32.125
of the situation.

49
00:02:32.515 --> 00:02:34.005
Basically what happens is

50
00:02:34.005 --> 00:02:37.205
that if you have probabilistic forecasts

51
00:02:37.265 --> 00:02:40.885
and you can sort of look at the, the near future, you say,

52
00:02:41.035 --> 00:02:44.885
okay, well, it looks like the, the air bounds

53
00:02:44.905 --> 00:02:46.725
around the forecast are small.

54
00:02:47.285 --> 00:02:51.405
I don't have to think about too much in reserve to, to,

55
00:02:51.665 --> 00:02:53.845
you know, fix up the problem if there is one.

56
00:02:53.995 --> 00:02:56.645
Whereas if they're wide, you might have

57
00:02:56.645 --> 00:02:58.845
to be flexible in the operation of

58
00:02:58.845 --> 00:02:59.845
The grid. So

59
00:02:59.845 --> 00:03:02.765
it helps the decision making on both sides.

60
00:03:02.985 --> 00:03:07.085
As far as I can see in the literature, there's a lot of,

61
00:03:07.515 --> 00:03:09.845
explanation of similar scenarios,

62
00:03:10.345 --> 00:03:13.205
but also there's a lot of different approaches

63
00:03:13.305 --> 00:03:16.125
to setting both deterministic

64
00:03:16.665 --> 00:03:19.885
and probabilistic, forecasts.

65
00:03:20.505 --> 00:03:23.995
And, the thing is that,

66
00:03:25.295 --> 00:03:29.435
my view is sometimes, and I don't wanna be too critical,

67
00:03:29.635 --> 00:03:32.195
'cause a lot of these papers are do very well in actually,

68
00:03:32.535 --> 00:03:34.875
you know, constructing these forecasts,

69
00:03:35.455 --> 00:03:38.315
but sometimes people I think,

70
00:03:40.025 --> 00:03:43.395
believe that they might have to use very powerful

71
00:03:44.085 --> 00:03:46.475
tools in order to do it properly,

72
00:03:47.135 --> 00:03:49.915
and they don't necessarily check

73
00:03:49.975 --> 00:03:53.835
to see if simple tools will work as well.

74
00:03:54.815 --> 00:03:58.875
Um, we've just run a recent experiment with, you know,

75
00:03:58.955 --> 00:04:01.675
a simple forecast from, solar

76
00:04:01.735 --> 00:04:05.955
and wind farms using very simple statistical models

77
00:04:06.135 --> 00:04:07.395
for a short time span.

78
00:04:07.655 --> 00:04:09.835
If you go a long time span, they won't work,

79
00:04:10.175 --> 00:04:11.595
but for a short time span,

80
00:04:12.065 --> 00:04:13.915
like the electricity market in Australia,

81
00:04:13.925 --> 00:04:16.115
quite often they will work quite well

82
00:04:16.375 --> 00:04:19.195
as some sophisticated machine learning tools,

83
00:04:19.895 --> 00:04:23.435
and yet they're f fully explainable in physical sense,

84
00:04:23.825 --> 00:04:27.475
whereas sometimes the machine learning ones, it's bit hard

85
00:04:27.475 --> 00:04:29.555
to unpack what they actually mean,

86
00:04:30.015 --> 00:04:32.715
and maybe they might be not transferable

87
00:04:32.735 --> 00:04:34.275
as easily and stuff like that.

88
00:04:34.495 --> 00:04:39.435
The literature, says a lot of good things about,

89
00:04:39.455 --> 00:04:40.555
different approaches,

90
00:04:40.815 --> 00:04:42.195
but the nice thing is that,

91
00:04:42.195 --> 00:04:45.915
especially a few recent papers talk about, you know,

92
00:04:46.135 --> 00:04:50.115
how useful it is in the sense that we've just explained

93
00:04:50.655 --> 00:04:52.035
for the generator

94
00:04:52.095 --> 00:04:54.795
and the, operator of the system as well.

95
00:04:55.615 --> 00:04:58.275
And I think that this leads into another little point

96
00:04:58.275 --> 00:05:00.195
that's quite interesting, I think,

97
00:05:00.815 --> 00:05:03.355
and that is that it's not just good enough

98
00:05:03.415 --> 00:05:06.875
to think about the skill of the forecasting method,

99
00:05:07.575 --> 00:05:11.555
but, how useful it is, is it actually of use.

100
00:05:12.055 --> 00:05:16.355
And in fact, you, researchers from the University

101
00:05:16.355 --> 00:05:19.035
of Alva Union in, Indian Ocean

102
00:05:19.055 --> 00:05:23.435
and myself have a new PhD student who has been working

103
00:05:23.535 --> 00:05:24.755
for the last year

104
00:05:24.755 --> 00:05:29.715
and over a year in Reunion Island, working on the value

105
00:05:29.775 --> 00:05:33.395
of forecasting rather than just the skill using,

106
00:05:33.425 --> 00:05:35.675
some examples from microgrids on that island.

107
00:05:36.375 --> 00:05:40.435
And, next year she'll be here with me working on,

108
00:05:40.435 --> 00:05:42.835
looking at the national electricity market

109
00:05:42.975 --> 00:05:47.035
and looking at the value of forecasting at it in terms of

110
00:05:47.575 --> 00:05:50.995
not just monetary means, but also so social means as well.

111
00:05:51.215 --> 00:05:53.955
So that was, a project funded

112
00:05:54.095 --> 00:05:56.675
by the Australian Renewable Energy Agency.

113
00:05:57.395 --> 00:05:58.405
Okay? Uh,

114
00:05:58.495 --> 00:06:02.325
after a request from the Australian energy market operator

115
00:06:02.585 --> 00:06:05.325
who run the grid and the market in Australia,

116
00:06:05.905 --> 00:06:08.925
or principally on the East coast over to South Australia

117
00:06:08.985 --> 00:06:11.525
and including Tasmania, it's too far to have,

118
00:06:11.525 --> 00:06:13.365
the grid extended to Western Australia.

119
00:06:13.365 --> 00:06:14.565
They operate their own.

120
00:06:15.625 --> 00:06:19.965
Um, now, and this was a project called for

121
00:06:20.025 --> 00:06:22.445
by Arena Renewable Energy Agency.

122
00:06:22.945 --> 00:06:27.565
Uh, I think there was about 12 consortia, working

123
00:06:27.995 --> 00:06:31.125
with different solar farms and wind farms to try

124
00:06:31.145 --> 00:06:34.925
and improve, the five minute forecasting

125
00:06:35.425 --> 00:06:37.725
of the output from those, farms.

126
00:06:38.305 --> 00:06:43.165
Now, the reason for that was that originally, the

127
00:06:43.905 --> 00:06:47.565
amo the operator itself did all that forecasting

128
00:06:48.065 --> 00:06:50.325
for all the different solar farms and wind farms,

129
00:06:50.425 --> 00:06:51.725
but then it became too many.

130
00:06:52.795 --> 00:06:53.855
So they said, okay,

131
00:06:53.855 --> 00:06:57.135
you've gotta do some self forecasting, okay?

132
00:06:57.395 --> 00:07:01.655
So we understand what might be available from these,

133
00:07:02.335 --> 00:07:06.405
variable renewable energy, the systems in the, in,

134
00:07:06.585 --> 00:07:09.445
you know, five minute, future and stuff like this.

135
00:07:10.225 --> 00:07:13.125
Um, trouble was that most of the solar

136
00:07:13.125 --> 00:07:16.405
and wind farms didn't actually have, people

137
00:07:16.975 --> 00:07:19.205
adept in forecasting methods.

138
00:07:19.985 --> 00:07:24.565
And so they, amo asked the, energy,

139
00:07:24.945 --> 00:07:28.605
the arena to fund some projects,

140
00:07:28.625 --> 00:07:31.245
and we were one of those, we had five different solar

141
00:07:31.375 --> 00:07:32.525
farms involved.

142
00:07:33.225 --> 00:07:36.205
And basically what happened was,

143
00:07:36.805 --> 00:07:39.165
I used my simple statistical tools.

144
00:07:39.865 --> 00:07:43.685
Um, some people from University of New South Wales used,

145
00:07:44.325 --> 00:07:45.845
satellite images, to try

146
00:07:45.865 --> 00:07:49.405
and predict what was going to, how the clouds are going

147
00:07:49.405 --> 00:07:51.085
to move, shall we say, and then,

148
00:07:51.185 --> 00:07:53.325
and evolve in the next five minutes.

149
00:07:53.985 --> 00:07:58.325
Um, some people from the C-S-I-R-O in Australia also had

150
00:07:58.985 --> 00:08:01.765
sky cameras installed at the solar farms,

151
00:08:01.945 --> 00:08:04.165
and they were looking at sky images

152
00:08:04.225 --> 00:08:05.525
and once again, trying

153
00:08:05.525 --> 00:08:08.045
to bring cloud movement, cloud evolution.

154
00:08:09.145 --> 00:08:13.045
And then we looked at using machine,

155
00:08:13.105 --> 00:08:16.365
simple machine learning tools to blend these forecasts

156
00:08:16.665 --> 00:08:19.485
to improve the overall efficiency of the forecast.

157
00:08:20.545 --> 00:08:23.565
Now, once again, the goal was not to

158
00:08:24.515 --> 00:08:27.885
improve the skill of the forecast, but to try

159
00:08:28.185 --> 00:08:33.055
and avert penalties for the solar farms from

160
00:08:33.055 --> 00:08:35.415
that cause or pay penalty, okay?

161
00:08:35.415 --> 00:08:36.695
Because that was the thing.

162
00:08:36.695 --> 00:08:40.535
They were incurring huge penalties,

163
00:08:41.355 --> 00:08:43.855
for mis forecasting,

164
00:08:43.915 --> 00:08:45.935
and they needed to improve that

165
00:08:45.995 --> 00:08:47.375
to improve their profitability.

166
00:08:47.795 --> 00:08:49.175
So that was the real goal.

167
00:08:49.555 --> 00:08:52.695
And we looked at specifically what

168
00:08:53.775 --> 00:08:54.855
combinations of

169
00:08:55.020 --> 00:08:58.405
Models seem to work better in particular situations.

170
00:08:59.145 --> 00:09:03.645
And so, that helped a EO,

171
00:09:04.105 --> 00:09:07.365
you know, do better in their Australian solar energy

172
00:09:07.645 --> 00:09:11.205
forecasting systems and Australian wind energy forecasting

173
00:09:11.205 --> 00:09:13.765
systems as well with the solar

174
00:09:13.825 --> 00:09:16.765
and wind farms, they don't at the present time

175
00:09:17.505 --> 00:09:22.165
bid into the system except just their expected forecast.

176
00:09:22.515 --> 00:09:27.165
Okay? That's it. Um, the more traditional,

177
00:09:27.505 --> 00:09:28.565
generators,

178
00:09:29.075 --> 00:09:32.445
they submit bids from minus a thousand dollars

179
00:09:32.545 --> 00:09:36.645
to plus I think $17,000 per megawatt,

180
00:09:36.785 --> 00:09:39.125
per megawatt, hour.

181
00:09:39.705 --> 00:09:42.485
And, that's,

182
00:09:43.425 --> 00:09:45.325
at diff 10 different price bands.

183
00:09:45.465 --> 00:09:47.365
And so they'll nominate at each price band

184
00:09:47.635 --> 00:09:49.085
what they can supply

185
00:09:49.785 --> 00:09:52.685
and, in the next five minutes, that sort of thing.

186
00:09:53.425 --> 00:09:57.605
Um, at the moment, solar

187
00:09:57.705 --> 00:10:00.805
and wind farms only give a particular,

188
00:10:01.465 --> 00:10:03.645
forecast, okay?

189
00:10:04.225 --> 00:10:06.405
Or they may not submit a forecast if they don't want

190
00:10:06.405 --> 00:10:09.685
to participate in that particular five minute interval.

191
00:10:10.665 --> 00:10:13.645
Um, this will possibly change,

192
00:10:13.745 --> 00:10:16.405
but I'm not sure exactly how in the future,

193
00:10:16.405 --> 00:10:20.805
because for instance, in South Australia where I am

194
00:10:21.485 --> 00:10:24.355
situated, the goal is,

195
00:10:24.455 --> 00:10:27.715
and it's just about to be put into legislation,

196
00:10:27.775 --> 00:10:32.435
as I understand that by 2027, a hundred percent

197
00:10:32.435 --> 00:10:34.475
of electricity will come from renewables.

198
00:10:36.155 --> 00:10:40.255
And that's, that's not every minute of the, of the year.

199
00:10:40.645 --> 00:10:43.415
Okay? Over the year, the total,

200
00:10:43.645 --> 00:10:46.775
generated will be a hundred percent of the supply.

201
00:10:47.555 --> 00:10:51.055
So sometimes the amount generated would be more than the,

202
00:10:51.515 --> 00:10:53.775
sorry, the demand,

203
00:10:53.965 --> 00:10:56.575
more than demand that'll be sent off to other states

204
00:10:57.035 --> 00:10:58.655
and vice versa at certain times.

205
00:10:59.035 --> 00:11:02.295
But over the year, the average amount should be equal,

206
00:11:02.295 --> 00:11:03.815
average amount should be equal.

207
00:11:04.395 --> 00:11:07.615
And so when that happens, of course, the,

208
00:11:07.675 --> 00:11:09.495
the idea at the moment is that the solar

209
00:11:09.495 --> 00:11:11.775
and wind farms are basically price takers.

210
00:11:12.485 --> 00:11:17.375
Okay? So the price is set by the traditional generators,

211
00:11:17.395 --> 00:11:19.575
but if there's no traditional generators, what do you do?

212
00:11:19.755 --> 00:11:22.055
You've gotta set some other type of mechanism.

213
00:11:22.155 --> 00:11:26.455
So probably at least there'll be some sort of,

214
00:11:28.675 --> 00:11:31.525
idea of how much you can supply at different,

215
00:11:31.695 --> 00:11:35.085
price levels, maybe, from the solar and wind farms.

216
00:11:35.185 --> 00:11:39.085
But in order to do that, that once again, that,

217
00:11:39.155 --> 00:11:43.325
probabilistic forecast will allow them to

218
00:11:43.875 --> 00:11:48.045
sort of decide the risk they can take at different levels.

219
00:11:48.325 --> 00:11:50.925
I mean, if you, if you think, okay, I'm,

220
00:11:51.345 --> 00:11:52.725
I'm gonna make sure that I

221
00:11:52.725 --> 00:11:55.885
Get into the market, so you might put a low value on your,

222
00:11:55.985 --> 00:11:59.605
the amount that you can, a low dollar value,

223
00:11:59.705 --> 00:12:01.525
for high amounts of electricity.

224
00:12:01.985 --> 00:12:03.765
But, you know, if you mis forecast,

225
00:12:03.785 --> 00:12:05.245
you're gonna get penalized a lot.

226
00:12:05.875 --> 00:12:09.005
Okay? Um, but if the price is high,

227
00:12:09.915 --> 00:12:11.525
that might be a risk worth taking.

228
00:12:12.355 --> 00:12:15.565
Okay? So, you know, it, it is probably like that.

229
00:12:15.845 --> 00:12:17.205
I mean, I'm speculating

230
00:12:17.205 --> 00:12:20.885
because even the, the operators don't know exactly

231
00:12:20.945 --> 00:12:22.965
how the market is going to change yet.

232
00:12:23.165 --> 00:12:26.125
I don't think. It doesn't look like in the, you know,

233
00:12:26.125 --> 00:12:29.165
the integrated system plan that's just been produced and,

234
00:12:29.465 --> 00:12:32.445
and other sorts of, documents from a emo

235
00:12:32.505 --> 00:12:35.085
and I think it's under discussion, to be honest,

236
00:12:35.825 --> 00:12:36.845
how this will operate.

237
00:12:36.905 --> 00:12:41.335
But I'm sure that whatever the market structure is,

238
00:12:41.345 --> 00:12:42.455
especially for solar

239
00:12:42.475 --> 00:12:44.335
and wind farms, that,

240
00:12:45.565 --> 00:12:49.215
putting air bounds on the forecast will help them make their

241
00:12:49.455 --> 00:12:54.095
decisions on the financial side of things better be one

242
00:12:54.095 --> 00:12:56.255
of the things we found out from it, well,

243
00:12:56.255 --> 00:12:59.575
it seemed like from the, projects we did was that,

244
00:12:59.875 --> 00:13:03.055
you know, there's different components to the companies

245
00:13:03.055 --> 00:13:04.735
that run these, farms.

246
00:13:05.195 --> 00:13:09.255
And one of them is how we actually manage all the financial,

247
00:13:09.755 --> 00:13:13.575
actions rather than the sort of technical actions of,

248
00:13:13.795 --> 00:13:15.415
you know, how you run the farm itself.

249
00:13:15.755 --> 00:13:18.135
How are you gonna make money out of it on the market?

250
00:13:18.675 --> 00:13:22.305
And so, you know, these people will be, you know,

251
00:13:22.455 --> 00:13:25.665
vitally looking at things like probabilistic forecasts if

252
00:13:25.665 --> 00:13:29.345
they don't do right now, the little experiments we've done

253
00:13:29.405 --> 00:13:33.745
so far at the moment for short-term forecasting seem

254
00:13:33.765 --> 00:13:35.865
to show that the,

255
00:13:37.345 --> 00:13:40.995
machine learning tools, at least the ones we've tried,

256
00:13:40.995 --> 00:13:42.195
simple neural networks

257
00:13:42.215 --> 00:13:45.955
and, long short-term memory systems

258
00:13:46.055 --> 00:13:50.115
and stuff like this seem to work, fairly

259
00:13:51.185 --> 00:13:54.995
similarly to simple, statistical models.

260
00:13:55.025 --> 00:13:57.395
When I say statistical models, I'm talking about,

261
00:13:57.665 --> 00:14:00.955
auto aggressive moving average plus some seasonality,

262
00:14:00.955 --> 00:14:03.195
especially with solar farms, obviously,

263
00:14:03.195 --> 00:14:05.555
because there's nothing generated at night

264
00:14:06.175 --> 00:14:07.875
and, all that sort of stuff.

265
00:14:08.495 --> 00:14:13.045
Uh, but, one, one place

266
00:14:13.045 --> 00:14:16.685
where we did, as I said, find that they were very useful was

267
00:14:16.685 --> 00:14:20.205
that idea of when you look at different approaches,

268
00:14:20.685 --> 00:14:24.925
physical models like, you know, sky cameras,

269
00:14:25.685 --> 00:14:28.525
NWP, neter, weather prediction models as well,

270
00:14:29.025 --> 00:14:30.805
and then, satellite images,

271
00:14:30.905 --> 00:14:33.245
and then maybe using statistical

272
00:14:33.245 --> 00:14:34.485
and everything like that as well.

273
00:14:34.505 --> 00:14:36.645
And then try the machine learning tools

274
00:14:37.205 --> 00:14:39.365
actually were worked really well at blending those

275
00:14:39.905 --> 00:14:43.205
and trying to figure out what, particular

276
00:14:44.485 --> 00:14:45.725
epochs shall we say,

277
00:14:45.785 --> 00:14:49.245
or particular times of the year, different types

278
00:14:49.305 --> 00:14:50.685
of models might work better.

279
00:14:50.785 --> 00:14:53.285
So you get different weights on those models.

280
00:14:53.465 --> 00:14:55.885
So, you know, possibly that's an area

281
00:14:55.885 --> 00:14:59.525
where they can be more useful than in a simple,

282
00:15:00.085 --> 00:15:01.205
straightforward forecast.

283
00:15:01.465 --> 00:15:04.895
One of the overriding things that you see in a lot

284
00:15:04.895 --> 00:15:09.835
of the literature is a concept for setting the,

285
00:15:10.835 --> 00:15:14.935
the, the distribution of the, of the bounds, shall we say,

286
00:15:15.355 --> 00:15:17.295
is a thing called quantile regression.

287
00:15:18.055 --> 00:15:22.035
Now, regression itself, most people have heard

288
00:15:22.035 --> 00:15:23.795
of things like line of best fit.

289
00:15:24.175 --> 00:15:26.875
So ordinary regression does that type of thing,

290
00:15:26.935 --> 00:15:30.515
and that's a, that basically forecasts the mean

291
00:15:30.975 --> 00:15:32.595
of the distribution, okay?

292
00:15:34.455 --> 00:15:36.975
Quantile regression, basically what you do is try

293
00:15:36.975 --> 00:15:40.615
to forecast the different percentiles of the distribution

294
00:15:40.675 --> 00:15:42.295
or quantiles, okay?

295
00:15:42.595 --> 00:15:45.055
And so you get this sort of all these,

296
00:15:45.315 --> 00:15:47.895
the whole distribution of the, of the errors

297
00:15:47.915 --> 00:15:49.895
as you go forward, all right?

298
00:15:50.595 --> 00:15:55.055
Now that seems to work pretty well,

299
00:15:55.355 --> 00:15:57.495
for a lot of, researchers.

300
00:15:57.995 --> 00:16:00.975
Uh, and it seems, quite a,

301
00:16:02.445 --> 00:16:06.205
a common method sometimes with some varieties to it,

302
00:16:06.555 --> 00:16:08.405
some improvements, et cetera, et cetera.

303
00:16:08.405 --> 00:16:10.765
But it seems to be a pretty prevalent method.

304
00:16:10.945 --> 00:16:14.725
So what I did in the most recent work was compare that

305
00:16:15.425 --> 00:16:18.195
to something that as far

306
00:16:18.195 --> 00:16:19.995
as I know nobody else has ever done,

307
00:16:20.655 --> 00:16:24.255
and that's to take the forecast

308
00:16:25.115 --> 00:16:30.085
for say, a five minute forecast, find the errors,

309
00:16:31.265 --> 00:16:36.215
and instead of doing quanta regression on those errors,

310
00:16:37.255 --> 00:16:38.655
actually take those errors

311
00:16:38.875 --> 00:16:43.255
and do some, transformation

312
00:16:43.255 --> 00:16:46.535
of the errors because they're not, they're not symmetric,

313
00:16:46.535 --> 00:16:48.615
they're certainly not normally distributed,

314
00:16:49.155 --> 00:16:51.815
but you can do a normalizing transformation

315
00:16:51.815 --> 00:16:54.615
of the heirs when you do that.

316
00:16:54.755 --> 00:16:58.455
You can use another method from financial mathematics

317
00:16:58.475 --> 00:17:00.935
called, arch

318
00:17:00.935 --> 00:17:03.735
and garch methods for forecasting the variance

319
00:17:05.335 --> 00:17:08.855
arch means auto aggressive conditional heteros castic,

320
00:17:09.485 --> 00:17:11.615
it's a mouthful, but see,

321
00:17:11.685 --> 00:17:15.575
something's heteros castic if the variance changes over

322
00:17:15.605 --> 00:17:18.975
time, which it does in these systems, okay?

323
00:17:19.205 --> 00:17:22.295
It's not constant like you'd like to have it, okay?

324
00:17:22.835 --> 00:17:25.175
You know, if you do a line of best fit, you're hoping

325
00:17:25.175 --> 00:17:27.455
that the variance around that line

326
00:17:27.525 --> 00:17:30.535
with the airs is constant, okay?

327
00:17:31.005 --> 00:17:33.495
With forecasting from solar or wind farms,

328
00:17:33.595 --> 00:17:36.175
or just the, the resources themselves,

329
00:17:37.235 --> 00:17:40.475
the variance changes over time, okay?

330
00:17:41.575 --> 00:17:45.675
Now, if you transform the errors to

331
00:17:46.255 --> 00:17:50.045
normal distribution, you can use these arch

332
00:17:50.065 --> 00:17:52.965
and garch methods to forecast the variance.

333
00:17:55.625 --> 00:17:59.765
Why you transform the normals that those methods rely on,

334
00:18:00.185 --> 00:18:03.855
the airs being normally distributed in order to do it.

335
00:18:04.195 --> 00:18:05.895
So you forecast the variance,

336
00:18:06.005 --> 00:18:09.815
then you can set these error bounds, you know,

337
00:18:10.135 --> 00:18:14.775
whatever level around the forecast you want back, transform

338
00:18:15.895 --> 00:18:20.555
to the, the distribution of the variance, of, sorry,

339
00:18:20.555 --> 00:18:22.955
of the distribution, of the errors that you originally had.

340
00:18:23.335 --> 00:18:25.195
And then you get the quantiles that way

341
00:18:25.825 --> 00:18:27.675
it's bit more sophisticated,

342
00:18:28.335 --> 00:18:32.995
but it seems to work better than the standard,

343
00:18:33.275 --> 00:18:35.075
quantile regression approaches.

344
00:18:35.855 --> 00:18:40.115
And, so it's just a little extra wrinkle that I, I,

345
00:18:40.385 --> 00:18:44.155
I've tried and it seems to improve the, the forecast,

346
00:18:44.215 --> 00:18:48.595
the probabilistic forecast when I say improve, okay?

347
00:18:48.975 --> 00:18:51.115
The, there's two things that you want to make sure

348
00:18:51.115 --> 00:18:53.235
that happen with the probabilistic forecast,

349
00:18:53.625 --> 00:18:55.435
that when you forecast forward,

350
00:18:56.405 --> 00:19:00.265
and if you check the actual afterwards, okay,

351
00:19:00.285 --> 00:19:03.945
if you're trying to make a 95% prediction interval from your

352
00:19:03.945 --> 00:19:07.705
probabilistic forecast, you hope that 95%

353
00:19:07.705 --> 00:19:10.465
of those observe values as they come along

354
00:19:11.425 --> 00:19:12.985
actually fall within that interval

355
00:19:13.365 --> 00:19:15.465
and, you know, other intervals as well.

356
00:19:16.005 --> 00:19:18.865
But also you want to have sharpness,

357
00:19:19.085 --> 00:19:23.305
so you want the interval as narrow as possible, okay?

358
00:19:23.805 --> 00:19:27.065
So in both cases, I got the right sort of coverage,

359
00:19:27.155 --> 00:19:31.865
shall we say, but with, with that transformation method,

360
00:19:32.125 --> 00:19:34.185
the intervals were a little bit sharper.

361
00:19:34.635 --> 00:19:37.905
Maybe there's a nice way where we can actually take, say,

362
00:19:38.325 --> 00:19:40.825
some, which I didn't, we didn't try

363
00:19:40.825 --> 00:19:44.545
with this last little experiment, take the forecast from,

364
00:19:45.005 --> 00:19:48.345
a simple statistical method

365
00:19:48.765 --> 00:19:52.145
and a forecast from the LSTM method

366
00:19:52.725 --> 00:19:55.105
and maybe figure out how to

367
00:19:56.245 --> 00:19:58.435
blend them in a nice optimal way.

368
00:19:59.105 --> 00:20:00.795
Because perhaps,

369
00:20:01.865 --> 00:20:05.115
even though the overall statistics are pretty similar

370
00:20:05.225 --> 00:20:09.275
with the two methods, maybe at certain types of situations,

371
00:20:09.855 --> 00:20:11.315
one works better than the other.

372
00:20:11.895 --> 00:20:14.075
And if you can, you know, somehow

373
00:20:14.855 --> 00:20:17.955
use some machine learning methods to, optimally

374
00:20:18.605 --> 00:20:21.475
blend those two, that could be really beneficial.

"Enhancing Energy Market Decisions: The Importance of Probabilistic Forecasting in Renewable Energy Generation"

00:00:00 - It's my belief that instead of just forecasting the expected, generation, shall we say, from a solar, wind farm in the next five minutes, which is the timeframe that the energy market works on in Australia, or even, you know, a longer period, it's better if you know the bounds around that forecast. I think are they, are they narrow? Is your forecast really sharp or are they wide? Is your forecast, you know, a bit loose, shall we say? because that can help the decision making both on, the generator's side and on the operator of the system
00:00:44 - I think with the, on the generator side, let's, have a little scenario. For instance, say you've got a, a solar farm and you have a battery, attached to it as well, which is becoming more and more common, or a wind farm as well. Okay? So you predict in the next five minutes, you're going to output a hundred megawatts from the, from the farm. Okay? Now, if the bands are, the bands around, or the, the possible error around that is small, okay, then you can be sort of more confident
00:01:25 - Because if you mis forecast, in other words, if your forecast is a hundred, but you produce 80 or 120, it may mean that the grid operator will actually penalize you for mis forecasting. There's a thing called cause or pays in the, energy market in Australia, for instance, if you under forecast, they have to bring more power on the grid from somewhere else. You basically have to pay for that fact. And if you, forecast too low, right? They may have to cut somebody else off, okay? That sort of thing. And so if your bands are wide
00:02:08 - I'd rather put that, electricity into my battery now and then send it to the grid when I know exactly what I can get out from the battery rather than put it into the grid and risk a penalty. Okay? But if they're narrow, the bands around the forecast, you're confident it's okay on the operator side of the, of the situation. Basically what happens is that if you have probabilistic forecasts and you can sort of look at the, the near future, you say, okay, well, it looks like the, the air bounds
00:02:47 - I don't have to think about too much in reserve to, to, you know, fix up the problem if there is one. Whereas if they're wide, you might have to be flexible in the operation of The grid. So it helps the decision making on both sides. As far as I can see in the literature, there's a lot of, explanation of similar scenarios, but also there's a lot of different approaches to setting both deterministic and probabilistic, forecasts. And, the thing is that,
00:03:29 - 'cause a lot of these papers are do very well in actually, you know, constructing these forecasts, but sometimes people I think, believe that they might have to use very powerful tools in order to do it properly, and they don't necessarily check to see if simple tools will work as well. we've just run a recent experiment with, you know, a simple forecast from, solar and wind farms using very simple statistical models

"Enhancing Short-Term Energy Forecasting: The Integration of Statistical Methods and Machine Learning in Australia's Renewable Energy Market"

00:04:07 - If you go a long time span, they won't work, but for a short time span, like the electricity market in Australia, quite often they will work quite well as some sophisticated machine learning tools, and yet they're f fully explainable in physical sense, whereas sometimes the machine learning ones, it's bit hard to unpack what they actually mean, and maybe they might be not transferable as easily and stuff like that. The literature, says a lot of good things about, different approaches, but the nice thing is that, especially a few recent papers talk about, you know,
00:04:50 - for the generator and the, operator of the system as well. And I think that this leads into another little point that's quite interesting, I think, and that is that it's not just good enough to think about the skill of the forecasting method, but, how useful it is, is it actually of use. And in fact, you, researchers from the University of Alva Union in, Indian Ocean and myself have a new PhD student who has been working for the last year and over a year in Reunion Island, working on the value
00:05:33 - some examples from microgrids on that island. And, next year she'll be here with me working on, looking at the national electricity market and looking at the value of forecasting at it in terms of not just monetary means, but also so social means as well. So that was, a project funded by the Australian Renewable Energy Agency. Okay? after a request from the Australian energy market operator who run the grid and the market in Australia, or principally on the East coast over to South Australia and including Tasmania, it's too far to have,
00:06:13 - They operate their own. now, and this was a project called for by Arena Renewable Energy Agency. I think there was about 12 consortia, working with different solar farms and wind farms to try and improve, the five minute forecasting of the output from those, farms. Now, the reason for that was that originally, the amo the operator itself did all that forecasting for all the different solar farms and wind farms, but then it became too many.
00:06:53 - you've gotta do some self forecasting, okay? So we understand what might be available from these, variable renewable energy, the systems in the, in, you know, five minute, future and stuff like this. trouble was that most of the solar and wind farms didn't actually have, people adept in forecasting methods. And so they, amo asked the, energy, the arena to fund some projects, and we were one of those, we had five different solar farms involved.
00:07:36 - I used my simple statistical tools. some people from University of New South Wales used, satellite images, to try and predict what was going to, how the clouds are going to move, shall we say, and then, and evolve in the next five minutes. some people from the C-S-I-R-O in Australia also had sky cameras installed at the solar farms, and they were looking at sky images and once again, trying to bring cloud movement, cloud evolution. And then we looked at using machine, simple machine learning tools to blend these forecasts

"Balancing Profitability and Forecasting Accuracy in Renewable Energy Markets"

00:08:20 - Now, once again, the goal was not to improve the skill of the forecast, but to try and avert penalties for the solar farms from that cause or pay penalty, okay? Because that was the thing. They were incurring huge penalties, for mis forecasting, and they needed to improve that to improve their profitability. So that was the real goal. And we looked at specifically what combinations of Models seem to work better in particular situations.
00:09:04 - you know, do better in their Australian solar energy forecasting systems and Australian wind energy forecasting systems as well with the solar and wind farms, they don't at the present time bid into the system except just their expected forecast. Okay? That's it. the more traditional, generators, they submit bids from minus a thousand dollars to plus I think $17,000 per megawatt, per megawatt, hour.
00:09:43 - at diff 10 different price bands. And so they'll nominate at each price band what they can supply and, in the next five minutes, that sort of thing. at the moment, solar and wind farms only give a particular, forecast, okay? Or they may not submit a forecast if they don't want to participate in that particular five minute interval. this will possibly change, but I'm not sure exactly how in the future, because for instance, in South Australia where I am
00:10:24 - and it's just about to be put into legislation, as I understand that by 2027, a hundred percent of electricity will come from renewables. And that's, that's not every minute of the, of the year. Okay? Over the year, the total, generated will be a hundred percent of the supply. So sometimes the amount generated would be more than the, sorry, the demand, more than demand that'll be sent off to other states and vice versa at certain times. But over the year, the average amount should be equal,
00:11:04 - And so when that happens, of course, the, the idea at the moment is that the solar and wind farms are basically price takers. Okay? So the price is set by the traditional generators, but if there's no traditional generators, what do you do? You've gotta set some other type of mechanism. So probably at least there'll be some sort of, idea of how much you can supply at different, price levels, maybe, from the solar and wind farms. But in order to do that, that once again, that, probabilistic forecast will allow them to
00:11:48 - I mean, if you, if you think, okay, I'm, I'm gonna make sure that I Get into the market, so you might put a low value on your, the amount that you can, a low dollar value, for high amounts of electricity. But, you know, if you mis forecast, you're gonna get penalized a lot. Okay? but if the price is high, that might be a risk worth taking. Okay? So, you know, it, it is probably like that. I mean, I'm speculating because even the, the operators don't know exactly how the market is going to change yet. I don't think. It doesn't look like in the, you know,

"Enhancing Financial Decision-Making for Renewable Energy through Probabilistic Forecasting Techniques"

00:12:29 - and other sorts of, documents from a emo and I think it's under discussion, to be honest, how this will operate. But I'm sure that whatever the market structure is, especially for solar and wind farms, that, putting air bounds on the forecast will help them make their decisions on the financial side of things better be one of the things we found out from it, well, it seemed like from the, projects we did was that, you know, there's different components to the companies that run these, farms. And one of them is how we actually manage all the financial,
00:13:13 - you know, how you run the farm itself. How are you gonna make money out of it on the market? And so, you know, these people will be, you know, vitally looking at things like probabilistic forecasts if they don't do right now, the little experiments we've done so far at the moment for short-term forecasting seem to show that the, machine learning tools, at least the ones we've tried, simple neural networks and, long short-term memory systems and stuff like this seem to work, fairly
00:13:55 - When I say statistical models, I'm talking about, auto aggressive moving average plus some seasonality, especially with solar farms, obviously, because there's nothing generated at night and, all that sort of stuff. but, one, one place where we did, as I said, find that they were very useful was that idea of when you look at different approaches, physical models like, you know, sky cameras, NWP, neter, weather prediction models as well, and then, satellite images, and then maybe using statistical
00:14:34 - And then try the machine learning tools actually were worked really well at blending those and trying to figure out what, particular epochs shall we say, or particular times of the year, different types of models might work better. So you get different weights on those models. So, you know, possibly that's an area where they can be more useful than in a simple, straightforward forecast. One of the overriding things that you see in a lot of the literature is a concept for setting the, the, the distribution of the, of the bounds, shall we say,
00:15:18 - Now, regression itself, most people have heard of things like line of best fit. So ordinary regression does that type of thing, and that's a, that basically forecasts the mean of the distribution, okay? Quantile regression, basically what you do is try to forecast the different percentiles of the distribution or quantiles, okay? And so you get this sort of all these, the whole distribution of the, of the errors as you go forward, all right? Now that seems to work pretty well,
00:15:57 - and it seems, quite a, a common method sometimes with some varieties to it, some improvements, et cetera, et cetera. But it seems to be a pretty prevalent method. So what I did in the most recent work was compare that to something that as far as I know nobody else has ever done, and that's to take the forecast for say, a five minute forecast, find the errors, and instead of doing quanta regression on those errors,

"Enhancing Forecast Accuracy through Variance Transformation and Model Blending Techniques"

00:16:38 - and do some, transformation of the errors because they're not, they're not symmetric, they're certainly not normally distributed, but you can do a normalizing transformation of the heirs when you do that. You can use another method from financial mathematics called, arch and garch methods for forecasting the variance arch means auto aggressive conditional heteros castic, it's a mouthful, but see, something's heteros castic if the variance changes over time, which it does in these systems, okay?
00:17:22 - You know, if you do a line of best fit, you're hoping that the variance around that line with the airs is constant, okay? With forecasting from solar or wind farms, or just the, the resources themselves, the variance changes over time, okay? Now, if you transform the errors to normal distribution, you can use these arch and garch methods to forecast the variance. Why you transform the normals that those methods rely on,
00:18:04 - So you forecast the variance, then you can set these error bounds, you know, whatever level around the forecast you want back, transform to the, the distribution of the variance, of, sorry, of the distribution, of the errors that you originally had. And then you get the quantiles that way it's bit more sophisticated, but it seems to work better than the standard, quantile regression approaches. And, so it's just a little extra wrinkle that I, I,
00:18:44 - the probabilistic forecast when I say improve, okay? The, there's two things that you want to make sure that happen with the probabilistic forecast, that when you forecast forward, and if you check the actual afterwards, okay, if you're trying to make a 95% prediction interval from your probabilistic forecast, you hope that 95% of those observe values as they come along actually fall within that interval and, you know, other intervals as well. But also you want to have sharpness, so you want the interval as narrow as possible, okay?
00:19:27 - shall we say, but with, with that transformation method, the intervals were a little bit sharper. Maybe there's a nice way where we can actually take, say, some, which I didn't, we didn't try with this last little experiment, take the forecast from, a simple statistical method and a forecast from the LSTM method and maybe figure out how to blend them in a nice optimal way. Because perhaps, even though the overall statistics are pretty similar
00:20:18 - one works better than the other. And if you can, you know, somehow use some machine learning methods to, optimally blend those two, that could be really beneficial.

Faculti AI can make mistakes

Research

Policy

Analysis