Optimal activity and battery scheduling algorithm using load and solar generation forecasts Yogesh Pipada Sunil Kumar

2025-05-02 0 0 661.05KB 6 页 10玖币

侵权投诉

Optimal activity and battery scheduling algorithm

using load and solar generation forecasts

Yogesh Pipada Sunil Kumar

University of Adelaide

Adelaide, Australia

yogeshpipada.sunilkumar,

Rui Yuan

University of Adelaide

Adelaide, Australia

r.yuan,

Nam Trong Dinh

University of Adelaide

Adelaide, Australia

trongnam.dinh,

S. Ali Pourmousavi

University of Adelaide

Adelaide, Australia

a.pourm@adelaide.edu.au

Abstract—Energy usage optimal scheduling has attracted great

attention in the power system community, where various method-

ologies have been proposed. However, in real-world applications,

the optimal scheduling problems require reliable energy fore-

casting, which is scarcely discussed as a joint solution to the

scheduling problem. The 5th IEEE Computational Intelligence

Society (IEEE-CIS) competition raised a practical problem of

decreasing the electricity bill by scheduling building activities,

where forecasting the solar energy generation and building

consumption is a necessity. To solve this problem, we propose

a technical sequence for tackling the solar PV and demand

forecast and optimal scheduling problems, where solar generation

prediction methods and an optimal university lectures scheduling

algorithm are proposed.

Index Terms—Forecasting, Reﬁned motif, Optimisation, Valley-

ﬁlling scheduling, Mixed-integer linear programming (MILP)

I. INTRODUCTION

A. Motivation

Even though infrequent, peak load is a cause for increased

capital and operating expenses for power networks. The reason

for this is twofold; the higher need for grid reinforcements

and the use of more expensive fossil fuel-based generators

to satisfy the peak loads (for a short duration). Thereby to

manage the extra costs due to peak demand, network operators

include a peak demand tariff (commensurate to the peak load)

in the electricity bills of commercial customers. This motivates

large-scale commercial customers such as universities and

manufacturers to manage their demand better and invest in

assets such as solar photovoltaic (PV) panels and stationary

batteries that have the potential to shift their demand and

reduce their electricity bills. A secondary effect of this is

the reduction in CO2emissions because of the lesser usage

of fossil fuel generators, which has the potential to combat

climate change and aid the process of decarbonization of our

power networks.

However, optimal scheduling of (schedulable) loads and

batteries (to minimize electricity costs) requires predictions

of inﬂexible load (baseload), solar PV generation and the spot

price of electricity. The ﬁrst two variables majorly depend on

weather conditions. Also, increasing renewable generation in

the generation mix has led to increasing price volatility in

the Australian national electricity market (NEM) which also

creates a correlation between electricity prices and the weather

[1]. This makes designing such algorithms challenging as most

commercial activities are planned over the mid to long term;

hence require reliable predictions. Additionally, the scheduling

problems are generally mixed-integer linear programs (MILPs)

which are NP-hard and may be intractable depending on the

formulation and problem size.

Surrounding this premise, the IEEE Computational Intel-

ligence Society (IEEE-CIS) partnered with Monash Univer-

sity (Victoria, Australia) to conduct a competition seeking

technical solutions to manage Monash’s micro-grid containing

rooftop solar panels and stationary batteries [2]. The main

challenge was to develop an optimal scheduling algorithm

for Monash’s lecture theatres and operation of batteries to

minimize their electricity bill, considering their baseload, solar

generation and NEM electricity spot prices. To this end, the

contestants were provided with actual time series data of the

building loads without any lecture program (baseload) and

solar generation from the micro-grid. So the contestants were

expected to predict the baseload and solar generation by taking

into account real-world weather data (also provided) for one

month in the future. Following this, the contestants had to use

actual electricity spot prices for the same duration along with

these predictions for the optimal scheduling algorithm.

B. Related work

We developed separate algorithms for the solar and baseload

forecasts based on practical insights and the given data. Many

approaches are available in the literature for both data types,

a relatively mature and crowded research ﬁeld. Therefore, the

methods we have developed for this problem borrow themes

from this mature literature.

For solar predictions, the basic theme used is similar to

the “clear sky” models for forecasting solar irradiance and

thereby estimating PV generation [3]. The general idea here

is to create a baseline model for PV generation, assuming

the sky is clear, meaning there is no cloud coverage or

temperature variance. This baseline is then modiﬁed based on

actual or expected weather conditions to estimate the actual PV

generation. The literature around this idea is based on physical

models of irradiance calculations, where equations are used

to develop the baseline for a given geographical location [3].

Newer methods use data-driven techniques such as time series

forecasting and machine learning models [4]. The data-driven

arXiv:2210.12990v1 [cs.LG] 24 Oct 2022

methods are gaining popularity because of their robustness,

speed and geographic adaptability.

Intuitively, this method gives reliable solar forecasts because

of solar generation’s seasonal and diurnal nature. However, the

main drawback is the lack of data speciﬁcally related to “clear

sky” days; Consequently, we use the most commonly occur-

ring day’s generation as baseline, which can be discovered by

our previous work [5].

For the baseload forecasting, the theme used is the appli-

cation of ensemble methods, as the current state of the art

identiﬁes these methods to be most accurate when compared

to stand-alone methods [6]. We use mainly a combination of

random forest (RF), gradient boost (GB), autoregressive inte-

grated moving average (ARIMA) and support vector machine

(SVM), which are all well-studied and standard methods for

time series forecasting [7]. Speciﬁcally, we apply different

forecasting methods to different sub-series after disaggregating

the load proﬁles using Seasonal and Trend decomposition

using Loess (STL) due to their cyclic patterns [8].

The optimal scheduling of distributed energy resources

(DER) and controllable/uncontrollable loads for micro-grids

comes under the class of problems commonly referred to

as energy management problems. However, these classes of

problems tend to be non-convex and non-linear in nature be-

cause of the constraints associated with the scheduled devices,

e.g., scheduling uninterruptible loads such as lectures that

cannot be stopped once started. As reviewed by the authors

in [9], the majority of these problems are modeled via classical

optimization approaches such as MILP [10] and mixed-integer

non-linear programs (MINLPs) [11]. Other approaches include

dynamic programming [12], rule-based optimization and meta-

heuristic algorithms [13].

Since these problems tend to be non-convex and non-linear,

the major challenge associated with solving these problems

is tractability and (in the case of meta-heuristic methods)

convergence to a global optimum. The mathematical literature

surrounding MILPs offers a variety of techniques to simplify

these problems before solving them, and the availability of

solvers such as Gurobi®for MILPs make them a very attractive

option for solving these problems. Also, unlike other models

where the algorithms (may) converge to a local optimal

solution, modern MILP solvers can obtain globally optimal

solutions for this class of problems.

C. Contributions

Based on the related work studied, this paper offers the

following contributions to the body of knowledge:

•A solar forecasting algorithm using training data from

reﬁned motif (RM) discovery technique and use of an

over-parameterized 1D-convoluted neural network (1D-

CNN) implemented via residual networks (ResNet). To

overcome the lack of “clear sky” data, we incorporated

the work done previously by Rui Yuan et al. [5] to identify

RMs in the given solar generation dataset. An RM is the

most repetitive pattern within a given time series, which

can be extracted along with the exogenous variables

(e.g., weather information) associated with this pattern.

Using this as a baseline, we estimated solar generation

by training an over-parameterized 1D-CNN. Some studies

have shown that over-parameterization of CNNs can lead

to better performance at the expense of longer training

time [14]–[16]. Therefore, a ResNet was implemented to

develop a deeper NN but with faster computation time.

•An optimal micro-grid scheduling algorithm is solved

based on real-world data for a university-based appli-

cation. To the best of our knowledge, there has not

yet been a study to co-optimize lectures schedule (and

associated resources) and battery operation (with PV

panels). Therefore, in this paper, we have developed and

tested our algorithm using real-world data and practical

case instances provided by Monash University. Given the

large problem size (one-month schedule) and the presence

of a quadratic term in the objective function for the cost of

peak demand, we proposed a two sub-problem approach

to formulate a tractable problem. The ﬁrst sub-problem

was used to limit the peak demand throughout the month,

eliminating the quadratic term from the objective. Then,

the peak demand was used to solve the second sub-

problem to minimize the total electricity costs.

The rest of the paper is structured as follows. Section II

introduces the data, information and problem requirements.

Section III proposes the forecasting and optimization method-

ologies. Section IV presents the numerical results of the

scheduling algorithm based on time series forecasting. Finally,

we conclude the paper section V.

II. BACKGROUND INFORMATION

This section describes the competition requirements and

data provided by the organizers.

A. Problem statement

This section describes the scheduling constraints and ob-

jective function provided by the competition organizers and

has been adapted from [2]. We were required to develop

prediction algorithms for the baseload of six buildings in

Monash University and the solar generation of PV panels

connected to them. After this, we had to use these predictions

to optimally schedule lecture activities and battery operation

while minimizing the electricity costs (i.e., electrical energy

consumption cost plus peak demand charge). We were allowed

to consider the electricity prices as known parameters to

simplify the problem and focus on predicting solar generation

and baseload.

For each lecture activity, we are provided with several small

or large rooms needed, the electrical power consumed per

room and the duration of the activities (in steps of 15 minutes).

We are also provided with a list of precedence activities, i.e.,

activities that must be performed at least one day before the

activity in question. For the batteries, we are provided with

the maximum energy rating, i.e., state of charge (SOC), the

peak charge/discharge power and charge/discharge efﬁciency.

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：

OptimalactivityandbatteryschedulingalgorithmusingloadandsolargenerationforecastsYogeshPipadaSunilKumarUniversityofAdelaideAdelaide,Australiayogeshpipada.sunilkumar,RuiYuanUniversityofAdelaideAdelaide,Australiar.yuan,NamTrongDinhUniversityofAdelaideAdelaide,Australiatrongnam.dinh,S.AliPourmousaviUniv...

展开>> 收起<<

Optimal activity and battery scheduling algorithm using load and solar generation forecasts Yogesh Pipada Sunil Kumar.pdf

共6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！

Optimal activity and battery scheduling algorithm using load and solar generation forecasts Yogesh Pipada Sunil Kumar

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: