Comparing the carbon costs and beneﬁts of low-resource solar nowcasting Ben Dixon

2025-04-27 0 0 1.77MB 8 页 10玖币

侵权投诉

Comparing the carbon costs and beneﬁts of

low-resource solar nowcasting

Ben Dixon∗

University College London (UK)

benjamin.dixon.21@ucl.ac.uk

Jacob Bieker

Open Climate Fix

jacob@openclimatefix.org

María Pérez-Ortiz

AI Centre, University College London (UK)

maria.perez@ucl.ac.uk

Abstract

Mitigating emissions in line with climate goals requires the rapid integration of low

carbon energy sources, such as solar photovoltaics (PV) into the electricity grid.

However, the energy produced from solar PV ﬂuctuates due to clouds obscuring

the sun’s energy. Solar PV yield nowcasting is used to help anticipate peaks and

troughs in demand to support grid integration. This paper compares multiple low-

resource approaches to nowcasting solar PV yield. To do so, we use a dataset of

UK satellite imagery and solar PV energy readings over a 1 to 4-hour time range.

Our work investigates the performance of multiple nowcasting models. The paper

also estimates the carbon emissions generated and averted by deploying models

such as these, and ﬁnds that short-term PV forecasting may have a beneﬁt several

orders of magnitude greater than its carbon cost and that this beneﬁt holds for small

models that could be deployable in low-resource settings around the globe.

1 Introduction

Solar photovoltaics (PV) is one of the safest and cleanest sources of energy, as measured in death

rates from accidents and air pollution, and greenhouse gas emission [11]. However, the ﬂuctuations

in the energy output, or yield, from solar PV makes it challenging for the electricity grid to rely at

large scale on solar energy. The yield of solar PV is inherently uncertain because clouds can obscure

the sun, and clouds can become thicker, thinner, and their movement across the sky is hard to predict

[2]. In order to guarantee stable energy supply, grid operators often have to keep gas spinning-reserve

online, releasing carbon dioxide into the atmosphere [7]. Forecasting over less than one day in

advance is called ‘nowcasting’ [13] to emphasize its dynamic use. Nowcasting solar energy input

into the grid accurately could reduce the need to keep the reserves on standby, allowing to use every

kilowatt efﬁciently and enhancing the grid resilience to the ﬂuctuations of renewable energy sources.

Since 2019, Open Climate Fix (OCF) has been working with National Grid ESO, the electricity system

operator for Great Britain, to build accurate PV nowcasting models that can enhance the resilience

of the energy grid [5]. This paper complements the work done by OCF. Speciﬁcally, we research

alternative low-resource methods (in terms of smaller/shallower neural networks architectures) for

predicting PV yield, and analyse their performance, their carbon footprints and aim to estimate, even

if roughly, the net emissions that could be averted by their potential use.

As more of the world incorporates solar PV into energy grids, there will be an increased need

to develop accurate solar PV forecasts. While a body of recent work has emphasized building

∗Primary author

Preprint. Under review.

arXiv:2210.04554v1 [cs.LG] 10 Oct 2022

more computationally efﬁcient models, most of the machine learning work still focuses on the

opposite: building larger models with more parameters to tackle more complex tasks [8]. Recent

work poses the question of how much is the performance gain worth [8], specially in these applications

where deployment of a simple technology could already have a big impact. The move towards less

computationally intensive models has already happened once in weather forecasting, as deep learning,

despite its competitiveness, has often lower computational requirements than numerical weather

prediction based on differential equations [12, 13].

Through our work, we study low-resource and shallower predictive models and try to estimate the

beneﬁt that they could bring if they were deployed large-scale for solar nowcasting

. The models,

which are build on one Tesla T4 GPU available through Google Colab, could be deployed large-scale

around the world independently of hardware limitations. The reason we are interested in researching

this low-resource setting, by studying smaller models, is because those are the models that could

be deployed to support the grid in emerging economies, which are where the real need is for clean

energy solutions, i.e. where the energy demand is rapidly growing.

2 An empirical comparison of PV yield nowcasting: Experimental results

For the task at hand, we study convolutional neural networks (CNNs), Long Short-Term Memory

networks (LSTMs) and their intersection [13] (ConvLSTMs). In theory, this comparison may seem

futile as LSTMs and ConvLSTMs should outperform CNNs, because of their use of state information.

However, Bai et al [1] argue that simpler CNNs can often outperform LSTMs. Moreover, ConvLSTMs

require a large amount of memory simultaneously available in order to process and update the state

information used in its predictions, which means they often take longer to train, leading to more

energy usage and hence higher emissions.

Dataset

Solar generation data refers to the yield from speciﬁc solar panels. The dataset

composed of satellite imaging and PV readings. The satellite images are provided by EUMETSAT,

covering the northern third of the Meteosat disc every ﬁve minutes. Open Climate Fix developed the

eumetsat_uk_hrv

dataset [6] which takes the ‘high resolution visible’ (HRV) channel for a reduced

time period and geospatial extent. The dataset was reduced in this paper to meet computing resources

available. A 64 x 64 crop was randomly chosen to focus on Devon, selecting images taken between

05:00 and 20:00 as there is minimal PV output outside of this window for most of the year. However,

this still left many readings in winter when the sun was below the horizon, and the PV yield was

zero, so readings when the solar altitude was below 10 degrees were also dropped. Open Climate Fix

provided a dataset of 1311 energy yield readings from PV systems in 5 minute increments for most

of 2018-2021, originally provided by a PV operator. The size of the dataset varied slightly based on

the prediction window. For the shortest window, there were 2018 observations of (12, 64, 64) and

(12, 1) to predict a series of 12 readings. For the longest prediction window, there were 659 blocks of

the same sizes described earlier to predict a series of 48 readings.

Learning task

We formulate our learning task as predicting a sequence of values representing

the future PV yield. All models take as input 12 sets of 5 min data, i.e., 1 hour of data as input,

and predict forward between 1 and 4 hours. We also consider two learning scenarios: i) learning

exclusively from past PV data only, and ii) learning from both past PV and satellite data as inputs.

Models compared

For each learning scenario (with and without satellite imaging), we test both

CNNs and LSTMs, our objective being comparing their prediction efﬁciency and carbon footprint at

the task at hand. Additionally, LSTMs encode time relationships explicitly, and we are interested

in evaluating if this additional complexity increases the prediction performance of our models to a

signiﬁcant extent (and at what potential carbon cost). For the PV only models (CNN and LSTM),

which are the simplest models we test, we take in past PV through multiple layers of CNNs, followed

by a fully connected layer. The other PV only model has a similar structure but uses LSTMs in place

of CNNs. The second group of models take both satellite images and past PV yield as input. The

Conv3D is a CNN which uses multiple levels to try to learn abstract features. After each convolutional

Our low-resource settings are not simulated, this project has been done as part of a MSc thesis dissertation,

with no other computational resources.

3Thanks to Open Climate Fix for providing the datasets.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Comparingthecarboncostsandbenetsoflow-resourcesolarnowcastingBenDixonUniversityCollegeLondon(UK)benjamin.dixon.21@ucl.ac.ukJacobBiekerOpenClimateFixjacob@openclimatefix.orgMaríaPérez-OrtizAICentre,UniversityCollegeLondon(UK)maria.perez@ucl.ac.ukAbstractMitigatingemissionsinlinewithclimategoalsrequ...

展开>> 收起<<

Comparing the carbon costs and beneﬁts of low-resource solar nowcasting Ben Dixon.pdf

共8页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Comparing the carbon costs and beneﬁts of low-resource solar nowcasting Ben Dixon

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: