Probabilistic Forecasting Methods for System-Level Electricity Load Forecasting Philipp Giese

2025-05-02 0 0 322.25KB 7 页 10玖币

侵权投诉

Probabilistic Forecasting Methods for System-Level

Electricity Load Forecasting

Philipp Giese

Technische Universit¨

at Berlin

Berlin, Germany

Philipp.Giese@campus.tu-berlin.de

Abstract—Load forecasts have become an integral part of

energy security. Due to the various inﬂuencing factors that can

be considered in such a forecast, there is also a wide range of

models that attempt to integrate these parameters into a system

in various ways. Due to the growing importance of probabilistic

load forecast models, different approaches are presented in this

analysis. The focus is on different models from the short-term

sector. After that, another model from the long-term sector is

presented. Then, the presented models are put in relation to

each other and examined with reference to advantages and

disadvantages. Afterwards, the presented papers are analyzed

with focus on their comparability to each other. Finally, an

outlook on further areas of development in the literature will

be discussed.

Index Terms—probabilistic load forecast, analyzing, short-

term, comparability,

I. INTRODUCTION

Many fundamental power system optimization problems

such as the unit commitment problem [1] take system load

and renewable generation as inputs [2] [3]. In recent years,

stochastic optimization [4], robust optimization, [5] and

distributionally robust optimization [6], [7] models have been

applied to solve the unit commitment problem. However, these

models require a probabilistic representation of uncertain

load and renewable generation, and thus deterministic,

point forecasts are not compatible with these optimization

models.Improving a prediction by as little as one percent can

reduce electrical costs by millions of dollars in a case of

10,000 MW [8].

On the other hand, the advances in machine learning models

have brought forth various successful deterministic forecasting

models for different power systems applications, including

load forecasting [9], PV forecasting [10], net load ramp

forecasting [11], and power system frequency forecasting

[12]. However, these models are primarily point forecasts,

which lack a comprehensive representation of the uncertainty.

The analysis of this work is structured as follows: In the

next section different models for the calculation of probabilis-

tic load forecasting from the literature are presented. Here,

the focus is especially on the basic functionality and less

on mathematical derivations. In this section, especially the

different approaches in the short-term sector are highlighted.

Afterwards, another model from the long-term sector is pre-

sented. In the following section, the problem of comparability

is discussed. Based on the article by T.Hong et at [8], the

previously presented papers are analyzed with focus on their

comparability. Finally, the results are summarized.

II. PROBILISTIC FORECASTING MODELS

The short-term methods aim to provide the most accurate

forecast possible for a brief period of time. Due to the

increasing number of available smart devices, it is becoming

increasingly difﬁcult to obtain precise loads forecasts based

only on weather data. Therefore, short-term forecast become

more important. In the following, different approaches are

shown which generate various forecast models because of

different regression models and machine learning. After that,

a probabilistic forecasting model is generated from the indi-

vidual models.

A. Combining Probabilistic Load Forecast

The approach described by Wang et al [16] aims to

generate different probabilistic load forecast models in the

ﬁrst step and to combine them in a common model in the

second step. For the generation of combined forecast models,

this approach mainly deals with 2 problems: On the one

hand, the goal is to generate different forecasts. For this

purpose, a large variance of features must be integrated to

allow inclusion of various uncertainty factors. For this step

different established quantile regression models are used. This

type of regression describes the data based on a probability

distribution: how likely is it that data is within a speciﬁc

range, so-called quantiles? To determine these, the article

cites several types. Neural networks and various types of

decision trees are given as examples. These will be trained

with the help of machine learning. The methods discussed in

the next section also use machine learning in different forms.

Most of the dataset is used to train the developed models.

In this section, the model is given the input parameters and

the output parameters to be achieved, from which the model

infers in which cases which decisions must be made. Based

on this, edge weights within the model are optimized so that

the input parameters lead to the desired output parameters.

The training database is divided into 4 parts. The ﬁrst 3

sections are used for individual model training and the last

section for combined model training. For individual training,

arXiv:2210.09399v1 [cs.LG] 17 Oct 2022

Fig. 1. Splitting of the training data sets [16]

the sections are used as follows: A part of the training data is

used for the adjustment of the single models generated by the

regression. The further data sets are used for the following

validation and testing of the model [16]. It is also possible to

reuse for validation the original training data in the form of

a sliding window [14].

The second phase of this system deals with the combination

of the generated models based on the corresponding model

weights. Here, the biggest problem is which combination is

the optimal one. For the combinations, different factors, such

as accuracy and uncertainty of the distribution, are considered.

To this purpose, error measuring functions, such as pinball lose

function, are used [16].

B. Feature Integration

1) Short Description: This approach, published by Chang

et al [13], generates suitable probabilistic load forecast models

based on the selection of a wide variety of data feature

combinations. A 2-stage method plus a subsequent test series

is applied. The main function of this model is that in the ﬁrst

stage of this 2-stage model all investigated point load forecast

models can be integrated into the model independent of their

own features. After that, a corresponding probabilistic model

can be generated from the point forecasts by selecting the

required features. In the stages 1 and 2 the system trains on the

one hand the calculation of the individual point load forecast

and on the other hand the recognition of the corresponding

features. Summarized, the individual point load forecast curves

result in the probabilistic intervals.

2) Functionality: In the ﬁrst stage, the individual models

for the calculation of point forecast are trained with historical

data. The models used were developed for various situations.

Accordingly, the models are based on different inputs and

features. Consequently, it is not guaranteed that the outputs

of the corresponding models have the same dimensions. In

this form the point load forecasts cannot be transformed into

probabilistic models. Therefore, a particularly important step

in this model is the detection and selection of the important

features or parameters of the individual models, which serve

as a foundation for the calculation of the probabilistic forecast

model. The selection of the features can be trained by using

Fig. 2. Functionality features interation [13]

different machine learning methods.

In the second stage, by selecting the features, the matching

point load forecast models are merged and together produce

the probabilistic forecast model. For the generation of point

load forecasts, the methodology of quantile regression and

error measuring is applied, similar to the approach used in the

previous section. The difference in this approach is that the

combination of individual forecasts is based on the selected

features. Just point forecast models with overlaps of the

selected features are used for generating the ﬁnal forecast.

The starting point of the testing are the input parameters,

from which the required features are identiﬁed to build the

probabilistic model. The system selects the corresponding

point load forecast models based on the selected features

and calculates the individual forecasts. The features and the

forecasts are combined to the probabilistic model. Finally, the

forecast is determined from this [13].

C. Clustering

1) Short Description: In the methodology by Wang [14]

the given data is grouped by using clustering algorithms into

similarity groups. Next, an individual forecast is calculated for

each cluster. The individual results were combined to a form

of probabilistic forecast.

2) Functionality: The model developed in this article can

be considered as a 3 phase model. In the clustering phase, the

individual data points are reduced to common dimensions or

parameters. Subsequently, the data is combined into clusters,

which are selected in such a way that they can be calculated

by suitable forecasting methods.

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：

ProbabilisticForecastingMethodsforSystem-LevelElectricityLoadForecastingPhilippGieseTechnischeUniversit¨atBerlinBerlin,GermanyPhilipp.Giese@campus.tu-berlin.deAbstractLoadforecastshavebecomeanintegralpartofenergysecurity.Duetothevariousinuencingfactorsthatcanbeconsideredinsuchaforecast,thereisalso...

展开>> 收起<<

Probabilistic Forecasting Methods for System-Level Electricity Load Forecasting Philipp Giese.pdf

共7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！

Probabilistic Forecasting Methods for System-Level Electricity Load Forecasting Philipp Giese

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: