EX-DSS: An Explorative Decision Support System for Designing and

Deploying Smart Plug Forecasting Pipelines

Giulia Rinaldi, Lola Botman, Oscar Mauricio Agudelo and Bart De Moor

KU Leuven, Department of Electrical Engineering (ESAT), STADIUS Center for Dynamical Systems, Signal Processing and

Data Analytics, Kasteelpark Arenberg 10, 3001 Leuven, Belgium

Keywords:

Decision Support System, Smart Plug Forecasting, Artiﬁcial Intelligence Pipeline.

Abstract:

Artiﬁcial Intelligence pipelines are increasingly used to address speciﬁc challenges, such as forecasting smart

plug loads. Smart plugs, which remotely control various appliances, can signiﬁcantly reduce energy consump-

tion in commercial buildings by about 20% when effectively scheduled using AI techniques. Designing these

AI pipelines involves numerous steps and variables, requiring collaboration and shared knowledge among de-

signers. A Decision Support System (DSS) can facilitate this process. This paper introduces the Explorative

Decision Support System (EX-DSS), which extends the classical DSS framework. The EX-DSS integrates an

Explorative Management Subsystem to provide project-speciﬁc recommendations and a Data Quality (DQ)

module to validate user inputs, ensuring clarity and enhancing information sharing. The EX-DSS architec-

ture framework was tested through a software prototype designed to create AI pipelines for forecasting smart

plug loads. The study found that using the EX-DSS improves the quality of suggestions, making them more

problem-speciﬁc and resulting in a more personalized and meaningful user experience, with a signiﬁcant po-

tential to reduce energy consumption in commercial buildings.

1 INTRODUCTION

Designing an Artiﬁcial Intelligence (AI) pipeline in-

volves multiple steps, from data cleaning to imple-

menting machine learning methods. This process can

be complex and time-consuming, especially when try-

ing to ﬁnd the most efﬁcient combination. In the con-

text of smart plug forecasting, this challenge is sig-

niﬁcant, as plug loads account for over 40% of total

energy consumption in commercial buildings, exclud-

ing lighting, HVAC, and water heating (Chia et al.,

2023). Smart plugs, which remotely monitor and

control electrical appliances, can save up to 20% of

electricity through effective scheduling, making AI

forecasting and scheduling methods promising solu-

tions (Botman et al., 2024).

A Decision Support System (DSS) can expedite

the design of these AI pipelines. A DSS is a ﬂexi-

ble software tool that assists in decision-making pro-

cesses and allows for shareability and reproducibility

among users, necessitating collaboration functionali-

ties and ways to evaluate user input.

This paper introduces an experimental Explorative

Decision Support System (EX-DSS) aimed at design-

ing and deploying industrial smart plug pipelines to

solve forecasting problems. The novel contributions

of this research include:

• Extending the conventional DSS with the Explo-

rative Management Subsystem, offering project-

speciﬁc insights and recommendations.

• Integrating a Data Quality (DQ) module to as-

sess the quality of user inputs, ensuring reliable

AI pipeline design.

• Maintaining a human-in-the-loop approach, giv-

ing users maximal control over every feature in

the EX-DSS.

• Providing a practical demonstration of the EX-

DSS through a software prototype for smart plug

forecasting, validating the system’s capabilities

and showcasing its effectiveness in real-world ap-

plications, thereby highlighting its potential im-

pact on reducing energy consumption in commer-

cial buildings.

This paper is structured as follows. Section 2

presents related works. In Section 3, the EX-DSS

architecture framework is introduced. Section 4 de-

scribes the application of this framework in the design

of smart plug forecasting AI pipelines via a software

prototype. In Section 5, results are discussed, and ﬁ-

Rinaldi, G., Botman, L., Agudelo, O. and De Moor, B.

EX-DSS: An Explorative Decision Support System for Designing and Deploying Smart Plug Forecasting Pipelines.

DOI: 10.5220/0012925800003838

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2024) - Volume 3: KMIS, pages 197-205

ISBN: 978-989-758-716-0; ISSN: 2184-3228

197

nally, in Section 6, conclusions are drawn, and future

research lines are presented.

2 BACKGROUND

A Decision Support System (DSS) is software that

helps users analyze data and make decisions. It gen-

erates insights and suggestions through a structured

framework (Gonzalez-Andujar, 2020). The classi-

cal DSS includes a Data Management Subsystem,

which stores and handles data; a Model Manage-

ment Subsystem, which manages models for DSS

tasks; a Knowledge Management Subsystem, which

provides information to users; and a User Interface,

which connects users with DSS subsystems (Turban

et al., 2010).

DSS supports semi-structured or unstructured de-

cisions, requiring evaluations beyond mathematical

modeling (Duan and Xu, 2009). For example, buy-

ing a new tool (Jacquet-Lagreze and Shakun, 1984)

or deciding on stock market entry/exit (Chandra et al.,

2007). These systems use rule-based models and

AI to predict scenarios and derive insights (Phillips-

Wren, 2013).

DSS applications are used in various ﬁelds like

business, healthcare, and agriculture. They can help

medical staff plan and monitor medications (Sloane

and J. Silva, 2020) and manage production costs in

agriculture (Rupnik et al., 2019). Tools like Rapid-

miner

and Knime

support the design of AI pipelines

but lack assistance during the conﬁguration phases.

The problem considered in the EX-DSS software

prototype involves forecasting the load of electrical

appliances, monitored and controlled by smart plugs

to optimize energy consumption (Chia et al., 2023).

Smart plugs can signiﬁcantly reduce energy usage in

commercial buildings by scheduling appliance oper-

ations. Accurate load forecasting is crucial for cre-

ating effective schedules. Current methods, including

time series analysis and AI-based techniques like neu-

ral networks and ensemble methods (Botman et al.,

2024), are time-consuming and complex to imple-

ment. This underscores the need for a robust DSS

framework like EX-DSS to streamline the process and

improve accuracy.

This paper introduces a knowledge-driven DSS

framework that incorporates detailed domain knowl-

edge. This makes it especially useful for complex

decision-making processes that require speciﬁc ex-

pertise. Speciﬁcally tailored for time series forecast-

https://altair.com/altair-rapidminer

https://www.knime.com/

ing, this framework focuses on input quality assess-

ment and fosters a collaborative research community

by enabling users to share insights, models, and re-

sults.

3 EX-DSS FRAMEWORK

As mentioned in Section 2, a traditional Decision

Support System (DSS) has three subsystems and a

User Interface (UI). The new Explorative Decision

Support System (EX-DSS) adds a Data Quality (DQ)

Module (Fig.1C) and an Explorative Management

Subsystem (Fig.1D.4). The DQ Module (Fig.1C) im-

proves data quality and promotes information sharing.

The Explorative Management Subsystem (Fig.1D.4)

provides project-speciﬁc insights for better analysis.

The EX-DSS architecture framework has two lev-

els: the User (Fig.1A) and the EX-DSS (Fig.1B). The

User interacts with the system to solve problems using

semi-automated steps. The EX-DSS analyzes prob-

lems, provides extra information, and helps create the

AI pipeline. A Graphical User Interface (Fig.1E) fa-

cilitates these interactions.

The EX-DSS has two main parts: the DQ

Module (Fig.1C) and the Pipeline Design Mod-

ule (Fig.1D). The DQ Module includes the Intake

Phase (Fig.1C.1) for uploading data and the As-

sessment Phase (Fig.1C.2) for evaluating data qual-

ity by generating a report. The Pipeline Design

Module (Fig.1D) has four subsystems: Data Man-

agement (Fig.1D.1), which stores essential informa-

tion; Model Management (Fig.1D.2), which man-

ages the pipeline structure (Fig.1D.2.a), conﬁguration

(Fig.1D.2.b) and training (Fig.1D.2.c); Knowledge

Management (Fig.1D.3), which provides general in-

formation; and Explorative Management (Fig.1D.4),

which derives project-speciﬁc insights from the key-

words extracted from the DQ report (Fig.1D.4.a).

The general knowledge in the Knowledge Man-

agement Subsystem (Fig.1D.3) is not tied to speciﬁc

problems. For example, it answers questions like

“What is forecasting?”. Project-speciﬁc knowledge in

the Explorative Management Subsystem (Fig.1D.4) is

detailed and tailored to speciﬁc projects, like “What

are the best methods for forecasting smart plugs?”.

The DSS is called “explorative” because it helps users

dive into and explore speciﬁc problems, such as fore-

casting smart plugs for energy savings in buildings.

By adding an extra subsystem, the EX-DSS im-

proves modularity, maintainability, and reusability. It

separates general knowledge from project-speciﬁc in-

formation, allowing better customization and system

performance.

KMIS 2024 - 16th International Conference on Knowledge Management and Information Systems

198

Interface (E)

Explorative Decision Support System (B)

Pipeline Design Module (D)

Explorative Management System (4)

Keyword Elaborator (a)

Knowledge

Management

System (3)

Model

Management

System (2)

Data Management

System (1)

Data Quality Module (C)

Guidelines

Assessment

(d)

Report

Generator (e)

Assessment Phase (2)Intake Phase (1)

Upload DATA (a)

Upload METADATA (b)

Upload PROBLEM

DESCRIPTION (c)

USER INPUT (1):

Data

Metadata

Probem Description

EX-DSS OUTPUT (2):

Assessment Report Pipeline Pipeline Results

USER (A)

USER FEEDBACK (3):

Block Choice

Configuration

Modification Requests

Human-in-the-loop

Suggestions

Modification

Requests

User

Feedback

User Input

Assessment

Report

Assessment

Report

User Input

Topological

Entity(a)

Hyperparameter

Entity(b)

Pipeline/

Pipeline Results

Deployment

Entity(c)

Pipeline/

Pipeline Results/

Suggestions

User

Feedback

Assessment

Report

User Input

Suggestions

Figure 1: EX-DSS architecture framework. This illustrates the proposed EX-DSS framework. (A) User-level input and

outputs. (B) Blocks composing the EX-DSS: (C) Block to assess the quality of the user’s input and (D) Block to aid the users

in designing, conﬁguring, and deploying the pipeline. (E) User Interface that allows the user to dialogue with the EX-DSS.

The arrows represent the ﬂow of information inside the systems.

4 EX-DSS FOR SMART PLUG

FORECASTING PIPELINES

This Section presents the application of the Explo-

rative Decision Support System (EX-DSS) frame-

work in the design of a smart plug forecasting AI

pipeline via a software prototype. Section 4.1 de-

scribes the use case and the context to implement

such EX-DSS, while Section 4.2 discusses in detail

the EX-DSS implementation.

4.1 The Context

The EX-DSS framework described in Section 3

addresses the design of Smart Plug Forecasting

pipelines. Smart plugs are devices that ﬁt between

appliances’ power cords and wall sockets, enabling

remote control. Smart plugs convert standard appli-

ances into smart ones. Various appliances like print-

ers, copiers, and TVs are monitored. Studies show

that plug loads (excluding lighting, HVAC, and water

heating) account for over 40% of total energy con-

sumption in commercial buildings, and that this con-

sumption is increasing over time (Chia et al., 2023;

Tuttle et al., 2020). Smart plugs can signiﬁcantly re-

duce energy consumption by automatically determin-

ing if a plug is idle or active and forecasting its usage.

This allows scheduling devices to turn on or off ac-

cordingly, leading to potential electricity savings of

up to 20%, optimizing building energy efﬁciency.

AI is crucial for determining plug usage, forecast-

ing consumption, and scheduling devices (Botman

et al., 2024). However, ﬁnding the best-performing

method is time-consuming. The software prototype

proposed in this paper aims to guide researchers in de-

signing pipelines for smart plug usage forecasting and

scheduling, thereby optimizing building energy con-

sumption. The trained model from the design phase

can also be downloaded for further use.

4.2 Description of the Software

Prototype

The EX-DSS software prototype is a web applica-

tion written in Python, implemented using the Dash

framework

. It assists users by providing descrip-

tions and suggestions for conﬁguring each block of

the AI pipeline, based on issue-speciﬁc keywords for

the Smart Plug problem described in Section 4.1 and

listed in the Appendix. Additionally, a chatbox al-

lows users to ask general questions. It is integrated

with the Natural Language Processing (NLP) model

developed by Cohere

https://dash.plotly.com/

https://cohere.com/

EX-DSS: An Explorative Decision Support System for Designing and Deploying Smart Plug Forecasting Pipelines

199

The functionalities of the EX-DSS software proto-

type are illustrated through an experiment using meth-

ods and datasets collected by Botman et al. (2024).

While Botman et al. (2024) proposed many alterna-

tives

, a subset of these methods is included in this

prototype. After conﬁguring the pipeline, the EX-

DSS connects to an external server for workﬂow man-

agement using the Airﬂow platform

4.2.1 New Dataset

The EX-DSS software prototype’s experiment began

with uploading a new dataset (Fig.1C.1). The user

provided a title, data source, and a short description

(Fig.1C.1.b) to facilitate understanding and improve

support in subsequent phases. Next, the user selected

keywords from a set proposed by the EX-DSS (see

Appendix for the full list). These keywords are cru-

cial for the Explorative Management Subsystem to

generate insights and target speciﬁc pipeline charac-

teristics. Finally, the user submitted the main data

ﬁle (Fig.1C.1.a) and could also upload supplemen-

tary ﬁles such as metadata or additional information

(Fig.1C.1.b).

For this experiment, the input was based on the

work proposed by Botman et al. (2024)::

• Title: Smart Plug Data,

• Data Source: “https://gitlab.esat.kuleuven.be/

Lola.Botman/smart-plug-pipeline/-/tree/main/

Dataset”

• Data Description: “The dataset is collected

through smart plug sockets between the wall plugs

and the electric appliance as detailed in Chia et al.

(2023). Smart plugs from Best Energy Reduc-

tion Technologies (BERT) are deployed in ﬁf-

teen buildings on the campus of the University of

California, San Diego (UCSD). The power level

of each appliance is recorded in mW at ﬁfteen-

minute intervals. The dataset consists of 169

high-quality smart plug time series spanning 498

days, from November 18

, 2021, to March 31

2023. The 169 plug loads include 146 printers, 16

copiers, 4 TVs, and 3 fax machines. This dataset

is openly accessible; see Botman et al. (2024) for

more details.”

• Keywords Selected: “Smart Plugs, High Perfor-

mance, Non-intrusive”

• Main File: “SmartPlugPreprocessedData.csv”

• Additional Metadata Files: [“SmartPlugMeta-

Data.csv” “SmartPlugHolidayData.csv”]

https://gitlab.esat.kuleuven.be/Lola.Botman/

smart-plug-pipeline

https://airﬂow.apache.org/

Table 1 shows a snapshot of the main dataset,

“SmartPlugPreprocessedData.csv.” Rows indicate the

time at which power values are recorded, and columns

represent monitored electrical appliances with power

values in mW. The goal is to predict these power val-

ues. In the illustrated subset, recordings start at 11:45

on the 18

of November 2021 until 17:00 on the 31

2023, capturing the power load of 169 smart plugs at

ﬁfteen-minute intervals.

4.2.2 Assessment Report

The Assessment Phase (Fig.1C.2) begins once the

system receives the data and the complementary in-

put. In this phase, the Data Quality module in EX-

DSS runs an internal analysis to assess the quality of

the inputs used in subsequent steps to generate sug-

gestions based on the work by Rinaldi. et al. (2023).

Although the Assessment Phase can be set up to

run different guidelines, for the software prototype,

three guidelines were chosen (Fig.1C.2.d):

• FAIR Paradigms (Wilkinson et al., 2016): The

system checks dataset uniqueness and corrects

saving (ﬁndability), attempts to open the dataset

in a pandas dataframe

(reusability), evaluates

data encoding (interoperability), and audited se-

curity rules (accessibility).

• Data Quality Analysis: This included check-

ing for missing values (completeness), assessing

value types (consistency), measuring distance be-

tween input and reference datasets (accuracy), and

verifying if the data is up-to-date (timeliness).

• Data Cleaning Automation: This involved rou-

tines like evaluating missing values, identifying

columns with a single value, and checking for row

duplication.

The clarity of the dataset description is also evalu-

ated. At the end of this phase, the system provides two

scores: one for data quality analysis and another for

data cleaning assessment, as shown in Rinaldi. et al.

(2023).

Although this analysis is done in the background,

users could conﬁgure it manually. They can choose

whether to analyze only the main ﬁle or include meta-

data ﬁles, select the weight of each quality metric, and

decide if any analysis should be skipped. All this

information is compiled into a report (Fig.1C.2.b),

which users can download.

Listing 1 shows a summary of the report that con-

tains the analysis with the above-mentioned guide-

lines.

https://pandas.pydata.org/

KMIS 2024 - 16th International Conference on Knowledge Management and Information Systems

200

Table 1: Smart Plugs Preprocessed Dataset snapshot. The dataset contains power loads of 169 smart plugs.

Timestamp Plug 0 Plug 1 Plug 2 ... Plug 167 Plug 168

2021-11-18 11:45:00 NaN NaN NaN ... NaN NaN

... ... ... ... ... ... ...

2021-11-18 13:15:00 NaN 10.413 19.656 ... NaN NaN

... ... ... ... ... ... ...

2023-03-31 17:00:00 55.809 10.03 19.44 ... 2.64 4.176

Listing 1: Extract of the JSON report containing the ﬁnd-

ings of the DSS after analyzing the information uploaded

by the user.

1 {”SCORES” {

2 ” D at a Q u a l i t y ” : 0 . 5 6 ,

3 ” D at a C l e a n i n g ” : 0 . 9 9 } ,

4 ”FAIR” {

5 ” R e u s a b i l i t y ” : ” The d a t a s e t was

r i g h t l y c o n v e r t e d t o a

d a t a f r a m e . ”

6 ” D a t a d e s c r i p t i o n c l a r i t y s c o r e ” :

” I n p u t t e x t c l a r i t y s c o r e :

85%” , . . . } ,

7 ”DATA QUALITY” {

8 ” T i m e l i n e s s ” : ” The d a t a s e t i s n o t

up t o d a t e −> p a r a m e t e r i s

0 . The u s e r s e t up t h e

t h r e s h o l d t o 0 y e a r s . ” . . . } ,

9 ”DATA CLEANING” {

10 ” Time Column ” : ” I d e n t i f i e d . ”

11 ” S i n g l e Columns ” : ” Th e r e a r e no

co lum ns w it h o n l y one v a l u e . ”

. . . }

4.2.3 New Project

The user needed to deﬁne the project’s title and de-

scription, upload any additional documents that could

help understand the project’s ﬁnal goal, and select the

most pertinent keywords from a subset proposed by

the EX-DSS (see Appendix for the full keywords list).

These keywords assist in formulating proper sugges-

tions and clarifying the research intention. An exam-

ple of a new project record is as follows:

• Title: Smart Plug Project,

• Project’s Description: “This project implements

smart plug active operating mode detection, plug-

level load forecasting, and a plug scheduling

methodology. A pipeline integrates the detected

operating modes with forecasting and scheduling,

aiming at reducing building energy consumption

(Botman et al., 2024)”

• Additional Documents: “No additional descrip-

tion documents”

• Keywords Selected: “High Performance, Non-

intrusive”

4.2.4 Pipeline Conﬁguration and the EX-DSS

Guidance

When a project is initialized, the user can start the de-

sign and conﬁguration of the pipeline. Figure 2 shows

a screenshot of the conﬁguration page of the EX-

DSS prototype software. Figure 2a lists the available

blocks developed in the EX-DSS software prototype,

while Figure 2b displays the suggestions generated by

the Knowledge Management Subsystem with support

from the Explorative Management Subsystem.

A suggestion consists of a block’s description, in-

sight generated by the Explorative Management for

the speciﬁc project, and a chatbox for communication

with the Knowledge Management Subsystem. The

Explorative Management Subsystem (Fig.1D.4) uses

the project’s keywords to ﬁnd the closest matching

dataset and generate suggestions based on stored ex-

pert knowledge. These suggestions include a list of

blocks and methods, such as the appropriate forecast-

ing method or technique to preprocess the dataset.

This modular approach provides accurate and special-

ized knowledge, which can be modiﬁed to enhance

speciﬁc subject knowledge or reused for other pur-

poses.

The Knowledge Management Subsystem

(Fig.1D.3) maintains a general knowledge view.

In the EX-DSS software prototype, it uses a Gen-

erative AI model developed by Cohere

. The

system connects to the Cohere platform via an API,

sends a request with a question, and displays the

AI-generated response to the user. Additionally, it

ensures the correct interconnections between blocks,

such as preventing the “Forecast” block from being

followed by the “Dataset” block.

Additionally, the Knowledge Management Sub-

system ensures that the interconnections between

blocks are correct. For example, in the EX-DSS soft-

ware prototype, the “Forecast” block could not be fol-

lowed by the “Dataset” block. Figure 2c represents

the conﬁguration area, where users choose methods

and conﬁgure parameters. For example, in the “Fore-

cast” block, users can select the forecasting method

and the prediction horizon. In the example shown

https://cohere.com/

EX-DSS: An Explorative Decision Support System for Designing and Deploying Smart Plug Forecasting Pipelines

201

(Fig. 2c), “Global XGBoost” was chosen with a pre-

diction horizon of “1” day.

Figure 2d shows the area where the user can orga-

nize and connect the chosen blocks. In the example,

the tasks were organized as follows: “Dataset” for se-

lecting the dataset, “Operating Mode” for determin-

ing the appliance’s operating mode, “Forecast” for

predicting the appliance loads, “Schedule” for plan-

ning when the appliances should be turned on or off,

and “Evaluate” for assessing the performance of the

other blocks. “Data Cleaning” and “Preprocessing”

were excluded since the dataset was already prepro-

cessed

Once the conﬁguration for each block is saved, the

pipeline is ready for training.

4.2.5 Pipeline Training and Results

Once the pipeline is ready, the user initiates the Model

Management Subsystem (Fig.1B.2), speciﬁcally the

Deployment Entity (Fig.1B.2.C), to launch the de-

ployment process. In the EX-DSS software proto-

type, this was implemented by an external server run-

ning Airﬂow

, a platform designed for executing and

monitoring a chain of tasks. The Deployment Entity

transferred the necessary ﬁles to the Airﬂow server.

Figure 3 displays the pipeline’s runtime, which took

9 minutes and 24 seconds to train and generate results.

Upon completion, the Airﬂow server communi-

cates the results back to the EX-DSS, and users can

download these results through the EX-DSS inter-

face. Additionally, users can download the forecast-

ing model, operating mode model, predictions, and ﬁ-

nal schedule. Table 2 provides an example of the ﬁnal

schedule. Additionally, Table 3 displays the numer-

ical values of the metrics used to assess smart plug

scheduling following the application of the “Global

XGBoost” method.

Listing 2: Example of static suggestion stored inside the

EX-DSS software prototype.

1{” P i p e l i n e S t e p s ” {

2” D at a C l e a n i n g ” ” I m p u t a t i o n ”

3” D at a C l e a n i n g ” ” N o r m a l i z a t i o n ”

. . . } ,

4” O p e r a t i n g Mode D e t e c t i o n ” {

5” Method ” ”GMM”

6” Method ” ” Ense mb le ”

7” F o r e c a s t i n g ” {

8” Method ” ” G l o b al XGBoost ”

9” Method ” ” G l o b al FNN”}

https://airﬂow.apache.org/

Table 2: Schedule Overview. The table provides an exam-

ple of the schedule for plugs 0 and 168. Both plugs are off

at night and turn on in the morning, with plug 0 turning on

earlier. Plug 0 turns off around lunch while plug 168 turns

off later in the afternoon. The schedule can be summarized

with the turn-on and turn-off times: Plug 0: {2023-01-13

05:15:00: ”Turn on”; 2023-01-13 12:30:00: ”Turn off”},

Plug 168: {2023-01-13 09:30:00: ”Turn on”; 2023-01-13

16:45:00: ”Turn off”}

Timestamp Plug 0 ... Plug 168

2023-01-12