Sustainable Energy Management System for AIoT Solutions Using

Multivariate and Multi-Step Battery State of Charge Forecasting

Farnaz Kasheﬁnishabouri

, Nizar Bouguila

and Zachary Patterson

Concordia Institute for Information Systems Engineering, Concordia University, 1515 St. Catherine W., Montreal, Canada

ﬁ

Keywords:

Artiﬁcial Intelligent of Things (AIoT), Responsible AI, Intelligent Energy Management,

Decision-Making System, Battery SoC, Time Series Forecasting, Renewable Energy Systems.

Abstract:

The convergence of Artiﬁcial Intelligence (AI) with Internet of Things (IoT) technologies, known as AIoT,

is revolutionizing industries, including smart cities. However, this transformation introduces challenges in

energy management. Addressing this issue while upholding responsible AI principles requires prioritizing

the sustainability of AIoT solutions through using renewable energy sources. While renewable energy offers

numerous advantages, its intermittent nature necessitates an effective power management system. Developing

a power management system serving as a decision-making platform for AIoT-driven solutions is the goal of

this study. This platform contains two critical components: accurate forecasts of battery “State of Charge”

(SoC), and the implementation of appropriate control strategies, including energy consumption adjustments.

This study focuses on accurate battery SoC forecasts, to this end, an experiment has been designed, and a data

logging system has been developed to produce suitable data since publicly available datasets do not match the

speciﬁc characteristics of this research. The SoC forecasting in this paper has been addressed as a multivariate

and multi-step time series forecasting problem, benchmarking various models. Comprehensive evaluations

on datasets with varying time intervals showed the Bi-GRU model outperforming others based on MAE and

RMSE metrics.

1 INTRODUCTION

In our increasingly interconnected world, the Inter-

net of Things (IoT) has emerged as a transformative

technology, changing how data is collected and used

across various domains. As a result of this widespread

adoption of IoT technology, many remarkable inno-

vations have taken place, particularly in the areas of

smart cities and healthcare. These innovations are,

however, accompanied by signiﬁcant challenges. The

extensive volume of data collected from different sen-

sors and devices during the advancement of IoT solu-

tions poses privacy issues that affect their widespread

adoption, as this data often contains sensitive infor-

mation (Lipford et al., 2022). The transmission of

large amounts of data from IoT devices requires a

high bandwidth, and not all environments support it

(Fetahu et al., 2022). Additionally, IoT data is often

unstructured, large-scale and unclean, which poses

challenges in extracting meaningful insights (Krish-

https://orcid.org/0000-0002-1661-9620

https://orcid.org/0000-0001-7224-7940

https://orcid.org/0000-0001-8878-7845

namurthi et al., 2020). On the other hand, a valuable

application of Artiﬁcial intelligence (AI) is its capa-

bility to uncover insights and opportunities through

data analysis. It is imperative to emphasize the prin-

ciples of Responsible AI when it comes to AI-driven

solutions, which include privacy, ethics, and sustain-

ability. Responsible AI ensures that AI technologies

are developed and used in ways that protect individ-

ual privacy, maintain ethical standards, and reduce

environmental impact by promoting energy-efﬁcient

practices and energy management (Barredo Arrieta

et al., 2020). AI integration with IoT ecosystems has

empowered them with the capability to analyze vast

amounts of data, derive insights, and facilitate au-

tonomous decision-making. The convergence of AI

and IoT technologies, known as AIoT, is introducing

a new era of smart and adaptive systems, accelerat-

ing innovation and increasing efﬁciency across a vari-

ety of industries (Zhang and Tao, 2021). Smart cities

are one of the sectors where AIoT is bringing signiﬁ-

cant innovation. In order to leverage the full power

of AIoT, it is important to address IoT challenges

and adhere to responsible AI principles, in which pri-

Kasheﬁnishabouri, F., Bouguila, N. and Patterson, Z.

Sustainable Energy Management System for AIoT Solutions Using Multivariate and Multi-Step Battery State of Charge Forecasting.

DOI: 10.5220/0012624800003714

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

In Proceedings of the 13th International Conference on Smart Cities and Green ICT Systems (SMARTGREENS 2024), pages 49-56

ISBN: 978-989-758-702-3; ISSN: 2184-4968

vacy, ethics, and sustainability are all taken into ac-

count. As a result, BusPas Inc., a Montreal-based

company, is offering a smart platform using AIoT in

order to tackle these challenges and meet the current

demand for smart mobility. Their idea involves re-

placing regular bus stop signs with intelligent and in-

terconnected displays named “SCiNe” which stands

for “Smart City Network”. Besides showing real-

time information regarding bus schedules, this de-

vice is equipped with several IoT sensors, including

a light sensor, environment sensors, an Infrared sen-

sor, microphones, speakers, and a ﬁsheye camera.

These sensors enable the intelligent display to col-

lect valuable data at bus stops, making them strate-

gic hubs for gathering information on passenger and

vehicle ﬂows, environmental conditions, trafﬁc pat-

terns, ridership, and more. This data contributes to

enhancing mobility and optimizing transportation ser-

vices. Despite this, what truly distinguishes this intel-

ligent display is its embedded computational capabil-

ities. Through the use of AI at the edge, this device

enables real-time data processing at bus stops while

preserving user privacy by only transmitting descrip-

tive information, not only safeguarding privacy but

also streamlining data transmission due to the reduced

data volume after processing. The advancement of

AI technologies has resulted in larger and more pow-

erful models capable of performing complex tasks.

However, these larger models are energy-intensive,

which poses a challenge. Furthermore, IoT technolo-

gies are developing rapidly, creating new challenges,

one of which is managing energy resources efﬁciently

(Guenfaf and Zafoune, 2023). This challenge is fur-

ther compounded when integrating AI with IoT in

AIoT-driven solutions. To address these challenges

and uphold the principles of responsible AI, it is im-

portant to ensure the sustainability of these solutions.

The use of renewable energy sources is a promising

approach to addressing this challenge. Following this

approach, SCiNe is powered by a lithium-ion battery

that derives its charge from a solar panel, showcas-

ing its dedication to sustainability. Renewable energy

sources, like solar panels, offers beneﬁts such as re-

ducing reliance on fossil fuels and cutting costs. It

can be harnessed in various locations, reducing de-

pendence on centralized grids and increasing energy

self-sufﬁciency. However, challenges exist in manag-

ing the intermittent nature of renewable sources for

a consistent energy supply for AIoT devices (Rathod

and Subramanian, 2022). Moreover, when renewable

energy sources are not producing enough power, en-

ergy storage is required to ensure a reliable energy

supply. However, current storage technologies have

capacity and efﬁciency limitations (Bharatee et al.,

2022). Therefore, an effective power management

system is essential to mitigate these challenges. In

systems based on renewable energy sources, predic-

tive control techniques have recently been offered

as a way to cope with uncertainty and intermittency

of energy production and consumption (Elmouatamid

et al., 2020). The State-of-Charge (SoC) of batteries,

which indicates the amount of energy stored in them,

is one of the main parameters for the development of

these techniques. Consequently, effective power man-

agement of renewable energy systems involves two

key aspects: accurately forecasting batteries SoC, and

the implementation of appropriate control strategies

such as adjusting energy consumption patterns, to en-

sure stable and reliable system operation. This power

management system serves as a decision-making sys-

tem, deﬁning different service levels for the device.

At each service level, certain functionalities of the de-

vice are limited to control power consumption. Based

on the predicted SoC of the battery, and considering

weather and solar conditions in the future, the sys-

tem switches between these service levels dynami-

cally. The main objective of this study is to develop

such an effective power management system for AIoT

solutions, exempliﬁed by our work on SCiNe, with

a speciﬁc focus on accurate battery SoC forecasting.

This SoC forecasting plays a central role within the

broader decision-making system, all while making ef-

ﬁcient use of renewable energies to ensure the practi-

cal sustainability of AIoT solutions. The majority of

previous work in battery SoC forecasting has focused

on microgrid systems, electric vehicles, and grid an-

cillary services. This study ﬁlls a gap in battery SoC

forecasting by focusing on a unique application do-

main: developing a power management system for a

battery-powered AIoT device charged through a so-

lar panel. Additionally, the characteristics and scale

of this project are different from previous work, so

publicly-available datasets are not applicable. The

remainder of this study is structured as follows: We

begin by presenting the state of the art in studies on

battery SoC forecasting in different domains and the

forecasting models employed for this purpose. Then,

the Design of Experiment section introduces our ex-

perimental design and data gathering process, empha-

sizing the custom data logging system developed to

collect relevant data. Finally, we exhibit the evalua-

tion results and ﬁnish with conclusions and avenues

for future research.

SMARTGREENS 2024 - 13th International Conference on Smart Cities and Green ICT Systems

2 LITERATURE REVIEW

Energy storage becomes necessary as a consequence

of using renewable energies when they fail to gener-

ate sufﬁcient power. Batteries are commonly used for

energy storage in Renewable Energy Source systems

and Lithium-ion batteries are often used in this con-

text. The SoC of batteries is one of the main param-

eters can be used in predictive control algorithms in

power management of systems using renewable en-

ergies. The process of determining the current state

of charge of a battery is known as SoC estimation,

whereas SoC forecasting involves predicting the fu-

ture state of charge based on past data and other fac-

tors. While numerous methods exist for SoC estima-

tion in batteries, limited research has been conducted

speciﬁcally on SoC forecasting which is an important

aspect of battery management systems (BMS) reliant

on renewable energy sources, as it enables the de-

velopment of effective power management strategies

(NaitMalek et al., 2021). Unlike other battery pa-

rameters, such as voltage, current, and temperature,

SoC cannot be directly measured. Researchers have

studied various methods for accurate SoC estima-

tion, which can be divided into three categories: con-

ventional methods, model-based methods, and data-

driven methods (Park et al., 2020). The focus of

this study is on SoC forecasting rather than estima-

tion. Several studies have investigated SoC forecast-

ing across different domains, employing various mod-

els and techniques. Researchers have used several

time series methods to forecast battery SoC in the

ﬁeld of grid ancillary services. For instance, (Ardian-

syah et al., 2021) developed an approach for multi-

step SoC forecasting of battery energy storage sys-

tems (BESS) in grid ancillary services, using Long

Short-Term Memory (LSTM) neural networks. This

model considers dynamic grid conditions and vary-

ing power demand, enabling accurate and reliable pre-

dictions of the battery SoC over multiple time steps.

(Ardiansyah et al., 2022) proposed a Seq2Seq regres-

sion approach for multivariate and multi-step fore-

casting of BESS in frequency regulation service. The

model showed promising results in accurately pre-

dicting SoC, enabling efﬁcient use of BESS. Another

domain in which SoC forecasting has been studied,

is in Micro grids, such as (NaitMalek et al., 2021).

That paper developed an embedded system for real-

time forecasting of battery SoC, enabling effective

energy management and decision-making in micro-

grid systems. In (Elmouatamid et al., 2020), MAP-

CAST, an adaptive control approach enhanced by pre-

dictive analytics, one of which is SoC forecasting,

is used to achieve energy balance in microgrid sys-

tems. By combining predictive analytics with adap-

tive control techniques, the proposed method opti-

mizes energy usage and ensures a stable energy bal-

ance, leading to improved energy management and

reliable operation in microgrid systems. One of the

key concerns of electric vehicle (EV) customers is

the driving range which depends mainly on battery

capacity. Forecasting battery SoC is therefore use-

ful in this context as well. In (NaitMalek et al.,

2022) Youssef NaitMalek et al. introduced a hybrid

method for accurate SoC forecasting in EVs. The ap-

proach combines a machine learning algorithm with

an EV model to forecast battery SoC. The machine

learning algorithm predicts vehicle speed, which is

then used as input for the EV model to determine

the battery SoC. The work presented in (NaitMalek

et al., 2019) also explored the integration of predic-

tive analytics techniques for multi-horizon forecast-

ing of battery SoC and contributes to the develop-

ment of an intelligent management system for battery-

powered electric vehicle. A variety of forecasting

techniques and algorithms have been applied in bat-

tery SoC forecasting. (NaitMalek et al., 2022) used

linear regression for SoC forecasting, which offered

simplicity and interpretability, making it suitable for

real-time SoC forecasting in battery-powered elec-

tric vehicles. However, to address the limitation of

not capturing complex nonlinear patterns in linear re-

gression, alternative algorithms such as decision trees

(DT) were employed in (Mashlakov et al., 2019). It is

worth noting that decision trees can be prone to over-

ﬁtting, and ensemble methods like random forests

(RF) or gradient boosting can be employed to mit-

igate this issue and further enhance SoC forecast-

ing performance. In light of this, (Mashlakov et al.,

2019) also applied random forest and Light Gradient

Boosting Machine (LightGBM), whereas (NaitMalek

et al., 2019) leveraged the power of Extreme Gradi-

ent Boosting (XGBoost). Moreover, in (Ardiansyah

et al., 2021), advanced deep learning architectures in-

cluding LSTM, Gated Recurrent Unit (GRU), Bidi-

rectional Long Short-Term Memory (Bi-LSTM), and

Bidirectional Gated Recurrent Unit (Bi-GRU) were

investigated. These recurrent neural network variants

demonstrated their ability to capture temporal depen-

dencies and long-term patterns in SoC data, enabling

more accurate and the multi-step forecasting of bat-

tery SoC. (Ardiansyah et al., 2022) proposed a so-

lution to the challenges of multi-step SoC forecast-

ing by using a sequence-to-sequence (seq2seq) model

in deep regression learning. This model has demon-

strated its effectiveness and robustness in various sce-

narios, making it a reliable approach for accurate mul-

tivariate and multi-step forecasting, as supported by

Sustainable Energy Management System for AIoT Solutions Using Multivariate and Multi-Step Battery State of Charge Forecasting

previous studies (Hewamalage et al., 2021). In sum-

mary, prior research on battery SoC forecasting has

been mostly centered around domains such as mi-

crogrid systems, electric vehicles, and grid ancillary

services. This research focuses on battery SoC fore-

casting which serves as the foundation for the intelli-

gent power management system designed to optimize

the operation of a battery-powered AIoT device us-

ing solar panel energy, addressing a crucial gap in the

ﬁeld of SoC forecasting - speciﬁcally for this unique

application. Considering the speciﬁc characteristics

and scale of this study, a custom dataset was required

and created since public data did not meet the study’s

needs. Three contributions were made in this paper:

the design of an experimental setup and the devel-

opment of a data logging system, handling the prob-

lem as a multi-step and multivariate time series fore-

casting, and conducting a comprehensive model eval-

uation. In order to overcome the shortage of suit-

able datasets in this domain, an experimental setup

was designed and a custom data logging system was

devised to ensure acquisition of relevant and accu-

rate data. Data collection is an essential part of the

study because it provides essential data for analysis

and model development. The paper uses various fore-

casting models, including machine learning and deep

learning approaches, to forecast SoC as a multivari-

ate and multi-step time series problem. Furthermore,

the paper conducts a comprehensive evaluation of the

forecasting models, providing valuable insights into

their performance.

3 DESIGN OF EXPERIMENTS

To gather data for this project, a comprehensive ex-

perimental setup was designed to capture and analyze

the relevant parameters of the battery system of the

device. This system comprised essential components,

including a lithium-ion battery, a solar panel (substi-

tuted with a programmable power supply for lab en-

vironment purposes), a load representing the device

with various subsystems, and an MPPT (Maximum

Power Point Tracking) solar charge controller. The

battery, power supply, and load were interconnected

through the MPPT charge controller, while a com-

puter served as the central control unit for data collec-

tion and control of the power supply. The computer

was connected to both the MPPT charge controller

and the power supply, allowing for real-time moni-

toring and control of the power supply’s parameters.

Figure 1 illustrates the setup conﬁguration. The

collected parameters from the MPPT charge con-

Figure 1: Setup Conﬁguration.

Table 1: Description of Parameters from MPPT Controller.

Category Parameters

Solar Info. Voltage, Current, State, Power

Battery Info. Voltage, Current, Temperature, SoC

Load Info. Current, Voltage, Power

Controller Info. Temperature, State

troller are presented in Table 1, with the data acqui-

sition facilitated by the MPPT controller’s software.

An experimental plan was designed to observe and

analyze the behavior of the battery system. The bat-

tery charging and discharging methodology were the

two crucial aspects that the experimental plan focused

on. Since the experiment was conducted in a con-

trolled lab environment without direct sunlight, a pro-

grammable power supply was used to charge the bat-

tery instead of solar panel. To simulate the charging

effect of the solar panel, solar radiation data for the

desired location were obtained from publicly avail-

able sources. Based on solar radiation values for each

hour, Equation 1 was used to calculate the voltage and

current settings for the power supply simulator. The

device’s solar panel speciﬁcations indicated a power

output of 50W under standard test conditions (STC)

with solar radiation at 1000W /m

. This information

was used to calculate the amount of power that the

solar panel would deliver to the battery based on the

actual solar radiation data.

Charge Power =

STC SR

× SP (1)

where: SR = Solar Radiation, STC SR = Solar Radi-

ation at STC, SP = Solar Panel Power at STC.

For instance, if the solar radiation for a particular hour

was measured as 300 W/m

, the power supply’s volt-

age and current were adjusted to deliver 15W to the

battery during that hour. This approach facilitated

the simulation of solar charging within the lab en-

vironment. One of the device’s functionalities is to

SMARTGREENS 2024 - 13th International Conference on Smart Cities and Green ICT Systems

send sensor data to the cloud at ﬁve-minute intervals

which is called telemetry data. This data is reﬂect-

ing the activation and deactivation of various subsys-

tems of the device. In this study, for the discharg-

ing methodology, the telemetry data for the month of

February 2023 was used. The collected telemetry data

were employed to precisely reproduce the activation

and deactivation patterns of the subsystems to simu-

late real-world scenarios throughout the experiment.

To ensure consistency in the data gathering process,

the experiment was initiated with a fully charged bat-

tery and incorporated a repetitive cycle. When the

battery SoC reached a critical level close to zero, the

experiment was temporarily paused. The battery was

then charged back to its full capacity, and the exper-

iment was resumed from the point at which the bat-

tery had reached the critical level. This process of

discharging, pausing, recharging, and resuming was

repeated multiple times throughout the experiment,

ensuring a reliable and controlled data collection ap-

proach. Through the carefully planned experiments,

which encompassed the simulated solar charging and

the replication of real-world subsystem activation and

deactivation, a comprehensive dataset was obtained.

This dataset serves as the foundation for the subse-

quent analysis and enables accurate battery SoC fore-

casting.

4 EXPERIMENTS

The dataset produced for this study consists of 40,320

observations, which corresponds to the number of

minutes in 28 days. The data was collected at one-

minute intervals to ensure a high level of detail in cap-

turing the behavior of the battery system. To assess

the impact of different time intervals on forecasting

accuracy, the dataset was resampled into three sub-

sets. This resampling makes a balance between data

granularity and computational efﬁciency, as the origi-

nal dataset contained a large volume of one-minute in-

terval observations. Based on the resampled datasets,

time series forecasting models were evaluated and re-

sults were compared across different temporal reso-

lutions. The subsets consisted of 2,688 observations

(15-minute interval), 1,344 observations (30-minute

interval), and 672 observations (1-hour interval). This

approach enables a comprehensive analysis of the

dataset and provides insight into the effect of tempo-

ral granularity on forecasting accuracy.

4.1 Data Preprocessing

An exploratory data analysis (EDA) is performed to

preprocess the dataset and select the important fea-

tures to include in the SoC forecasting model. In

Figure 2, the correlation heatmap, which is based on

Pearson correlation coefﬁcients, clearly shows that

the battery voltage and SoC are positively correlated,

with a correlation coefﬁcient of 1. This high correla-

tion is not surprising, considering that the MPPT con-

troller relies on the battery voltage as a crucial factor

in determining the SoC. The pearson correlation co-

efﬁcient r is calculated using Equation 2:

r =

· S

(2)

where S

represents the covariance between variables

X and Y, S

is the variance of X and S

is the vari-

ance of Y.

Figure 2: Pearson correlation heatmap displaying the rela-

tionship between the gathered features and the SoC of the

battery.

According to the correlation values, the SoC is

more correlated with load voltage, solar voltage, and

load current. Therefore, these features become cru-

cial factors to consider in order to improve the SoC

forecasting model. To further improve the feature se-

lection process, mutual information (MI) is employed

alongside correlation analysis. While correlation fo-

cuses on linear relationships between variables, MI

takes into account both linear and nonlinear depen-

dencies. It quantiﬁes the amount of information one

variable provides about another, capturing a broader

range of relationships beyond what correlation alone

can reveal. The MI is calculated as:

MI(X,Y ) =

∑∑

P(X,Y )log



P(X,Y )

P(X)· P(Y )



(3)

where P(X ,Y ) represents the joint probability distri-

bution of variables X and Y , and P(X) and P(Y ) rep-

Sustainable Energy Management System for AIoT Solutions Using Multivariate and Multi-Step Battery State of Charge Forecasting

resent their marginal probability distributions (Zhou

et al., 2022). Normalized mutual information (NMI)

is used to standardize evaluations across feature

scales. A high NMI score indicates strong depen-

dence between the target feature and the input feature.

NMI scores range from 0 to 1, with 1 representing

perfect correlation (0 = no mutual information). Fig-

ure 3 shows NMI scores between all the features and

battery SoC. Battery voltage, solar voltage, load volt-

age, and battery current have the highest NMI scores,

all with NMI score above 0.2. Therefore, based on

both analyses, the following features can be consid-

ered as features for improving the SoC forecasting

model: battery voltage, load voltage, solar voltage,

load current, and battery current.

Figure 3: Normalized mutual information score between

SoC and the features.

In the experiment, the Date and Time were set as

the time series index. The SoC percentage and the

features selected based on the Pearson correlation and

NMI score were used as input data.

4.2 Forecasting Models

Various multi-step and multivariate time series fore-

casting models were applied to the resampled datasets

with different time intervals to forecast the SoC of

the battery. Both machine learning and deep learn-

ing modeling approaches were used in this study.

Initial benchmarks were established using machine

learning models such as DT and RF. The data was

then modeled using deep learning models including

CNN, LSTM, GRU, Bi-LSTM, and Bi-GRU, in or-

der to capture complex temporal patterns. Forecast

horizons of 2 hours, 5 hours, and 10 hours were

used to evaluate the effectiveness of these models.

Training and testing of the models were conducted

on resampled subsets of data for each forecast hori-

zon. Using this approach, we were able to evaluate

model performance over various time intervals and

forecast horizons. The look-back window determines

the amount of historical data used for forecasting fu-

ture time steps. In this study, the look-back windows

are set to the same duration as the forecast horizons

in each model. Adam optimizer is used for tuning

the developed deep learning models, and the dropout

regularization to avoid overﬁtting. The forecasting

models were implemented using Keras 2.6.0 API with

Tensorﬂow 2.12.0 as the backend, within the Python

3.11.4 environment. Table 2 presents the parameters

used for training and evaluating the forecasting mod-

els,

Table 2: Parameters Set for the Models.

Parameter Value

Training and Testing portion 85% and 15% of total data

Validation portion 30% of training data

Normalization MinMax normalization (range 0 to 1)

Regularization Dropout = 0.2 in each layer

Early stop to avoid overﬁtting Patience = 30

Look-back window size range 2-40 data points

Optimization algorithm Adam

Maximum number of epochs 150

4.3 Evaluation

The forecasting models were assessed using mean

absolute error (MAE) and root mean squared error

(RMSE) to evaluate the accuracy of SoC forecasts.

MAE measures the average magnitude of prediction

errors. It is calculated as the mean of the absolute dif-

ferences between the predicted values ( ˆy

) and the ac-

tual values (y

) for each observation in the dataset, as

shown in Equation 4. In this equation, n is the number

of observations.

MAE =

∑

i=1

− ˆy

| (4)

RMSE is the square root of the average of the squared

differences between the predicted values ( ˆy

) and the

actual values (y

) for each observation in the dataset,

represented by Equation 5.

RMSE =

∑

i=1

− ˆy

)

(5)

Both MAE and RMSE express average model predic-

tion error in units of the variable of interest, which, in

this study, is the SoC of the battery. The SoC is ex-

pressed as percentage with 100% representing a fully

charged battery and 0% indicating an empty battery.

SMARTGREENS 2024 - 13th International Conference on Smart Cities and Green ICT Systems

Figure 4 provides a detailed comparison of the

forecasting models used in this study, based on the

MAE and RMSE performance metrics. Three resam-

pled datasets with 15-minute, 30-minute, and 1-hour

scales are used to evaluate the forecasting models,

covering forecast horizons of 2 hours, 5 hours, and 10

hours. The ﬁgure displays the errors for each model’s

last time step forecast, providing insights into the ef-

fectiveness of the models for long-term forecasting.

It is observed that RMSE errors are generally higher

than MAE errors, indicating RMSE’s sensitivity to

large errors. Additionally, and as expected, errors for

all models tend to increase with increasing forecast

horizons, reﬂecting the complexity and uncertainty

of long-term forecasts. The Bi-GRU model outper-

forms other models across various forecast horizons

and dataset scales, making it a good choice for more

accurate predictions.

Figure 4: Comparison of forecasting models for the last

time step on each dataset.

Further quantative evaluation of forecasting mod-

els was conducted in addition to visual analysis from

the ﬁgure. While the ﬁgure enabled comparison of

forecasts for the last time step, Table 3 provided an

assessment of the models’ overall performance. This

evaluation involved comparing the average errors for

all time steps across the longest forecast horizon of 10

hours. In this way, the assessment allowed for eval-

uating the model’s accuracy and reliability across the

entire forecast period, gaining a better understanding

of their overall performance and forecasting capabili-

ties. Bi-GRU model has the lowest MAE and RMSE,

indicating the highest accuracy among all models.

Table 3: Overall Forecast Performance on Test Data for the

Horizon of 10 Hours.

Models 15 min dataset 30 min dataset 1 hour dataset

MAE RMSE MAE RMSE MAE RMSE

DT 6.02 7.33 4.76 6.91 6.20 7.46

RF 4.49 5.71 3.56 5.36 4.70 5.84

CNN 4.45 5.60 3.91 4.92 4.39 5.59

LSTM 4.63 5.34 3.88 4.69 4.24 5.46

GRU 4.46 5.30 3.71 4.55 3.39 4.34

Bi-LSTM 3.88 4.69 3.70 3.42 2.93 3.72

Bi-GRU 3.61 4.57 2.55 3.26 2.79 3.66

Considering both visual and quantitative assess-

ments, the Bi-GRU model consistently outperforms

other models, highlighting the effectiveness of its

bidirectional architecture and gated recurrent units for

Battery SoC forecasting.

5 CONCLUSION

The integration of AI and IoT in AIoT-driven solu-

tions presents innovation, but also energy challenges.

To tackle these challenges while adhering to the prin-

ciples of responsible AI, it is crucial to prioritize the

sustainability of these solutions through the promis-

ing adoption of renewable energy sources. Despite

the beneﬁts of renewable energy, challenges such as

its intermittent nature necessitate the implementation

of an effective power management system. This study

focused on developing an effective power manage-

ment system, serving as a decision-making frame-

work for AIoT solutions. It was exempliﬁed by

the development of a battery-powered AIoT device

charged through a solar panel named “SCiNe”, with a

speciﬁc emphasis on accurate battery State of Charge

(SoC) forecasting. To understand the behavior of

system, an experiment was designed and a custom

data logging system was developed to gather relevant

data, enabling accurate analysis and model develop-

ment. The study explored the multivariate and multi-

Sustainable Energy Management System for AIoT Solutions Using Multivariate and Multi-Step Battery State of Charge Forecasting

step time series forecasting domain, using a variety of

models, including DT and RF, to deep learning mod-

els of CNN, LSTM, GRU, Bi-LSTM, and Bi-GRU.

The models were evaluated using both last time step

forecasts for a comparative view and average errors

over the entire forecast period for a comprehensive

evaluation. The Bi-GRU model outperformed other

models across datasets with varying time intervals and

forecast horizons. These ﬁndings highlight the poten-

tial of the Bi-GRU model for real-world applications

in similar systems. Incorporating additional input fea-

tures, such as weather and solar data, exploring alter-

native time series forecasting models, and integrating

the SoC forecasting solution into the decision-making

system, offer promising avenues for future enhance-

ments in the study. This study has primarily ad-

dressed the ﬁrst phase of the decision-making system

for managing AIoT device power – accurate battery

SoC forecasting. The next step is to design and im-

plement control strategies that enable dynamic adjust-

ments to service levels of the device. These service

levels deﬁne speciﬁc operating modes for the device,

with each level corresponding to different function-

alities and power consumption limits, ensuring both

system stability and power efﬁciency. These enhance-

ments lead to the development of a sustainable power

management system for AIoT applications.

REFERENCES

Ardiansyah, Kim, Y., and Choi, D. (2021). Lstm-based

multi-step soc forecasting of battery energy storage

in grid ancillary services. In 2021 IEEE Interna-

tional Conference on Communications, Control, and

Computing Technologies for Smart Grids (SmartGrid-

Comm), pages 276–281.

Ardiansyah, Masood, Z., Choi, D., and Choi, Y. (2022).

Seq2seq regression learning-based multivariate and

multistep soc forecasting of bess in frequency regula-

tion service. Sustainable Energy, Grids and Networks,

32:100939.

Barredo Arrieta, A., D

ıaz-Rodr

ıguez, N., Del Ser, J., Ben-

netot, A., Tabik, S., Barbado, A., Garcia, S., Gil-

Lopez, S., Molina, D., Benjamins, R., Chatila, R.,

and Herrera, F. (2020). Explainable artiﬁcial intelli-

gence (xai): Concepts, taxonomies, opportunities and

challenges toward responsible ai. Information Fusion,

58:82–115.

Bharatee, A., Ray, P. K., Subudhi, B., and Ghosh, A. (2022).

Power management strategies in a hybrid energy stor-

age system integrated ac/dc microgrid: A review. En-

ergies.

Elmouatamid, A., Ouladsine, R., Bakhouya, M.,

EL KAMOUN, N., Zine-dine, K., and Khaidar,

M. (2020). Mapcast: an adaptive control approach

using predictive analytics for energy balance in micro-

grid systems. International Journal of Renewable

Energy Research, 10:945.

Fetahu, L., Maraj, A., and Havolli, A. (2022). Internet

of things (iot) beneﬁts, future perspective, and im-

plementation challenges. 2022 45th Jubilee Inter-

national Convention on Information, Communication

and Electronic Technology (MIPRO), pages 399–404.

Guenfaf, Y. and Zafoune, Y. (2023). An iot ml-based system

for energy efﬁciency in smart homes. In 2023 IEEE

World AI IoT Congress (AIIoT), pages 0198–0203.

Hewamalage, H., Bergmeir, C., and Bandara, K. (2021).

Recurrent neural networks for time series forecast-

ing: Current status and future directions. International

Journal of Forecasting, 37(1):388–427.

Krishnamurthi, R., Kumar, A., Gopinathan, D., Nayyar, A.,

and Qureshi, B. (2020). An overview of iot sensor data

processing, fusion, and analysis techniques. Sensors,

20(21).

Lipford, H. R., Tabassum, M., Bahirat, P., Yao, Y., and Kni-

jnenburg, B. P. (2022). Privacy and the Internet of

Things, pages 233–264. Springer International Pub-

lishing, Cham.

Mashlakov, A., Honkapuro, S., Tikka, V., Kaarna, A., and

Lensu, L. (2019). Multi-timescale forecasting of bat-

tery energy storage state-of-charge under frequency

containment reserve for normal operation. In 2019

16th International Conference on the European En-

ergy Market (EEM), pages 1–8.

NaitMalek, Y., Najib, M., Bakhouya, M., and Essaaidi, M.

(2019). On the use of machine learning for state-of-

charge forecasting in electric vehicles. In 2019 IEEE

International Smart Cities Conference (ISC2), pages

408–413.

NaitMalek, Y., Najib, M., Bakhouya, M., and Essaaidi, M.

(2021). Embedded real-time battery state-of-charge

forecasting in micro-grid systems. Ecological Com-

plexity, 45:100903.

NaitMalek, Y., Najib, M., Lahlou, A., Bakhouya, M.,

Gaber, J., and Essaaidi, M. (2022). A hybrid approach

for state-of-charge forecasting in battery-powered

electric vehicles. Sustainability, 14(16):9993.

Park, S., Ahn, J., Kang, T., Park, S., Kim, Y., Cho, I., and

Kim, J. (2020). Review of state-of-the-art battery state

estimation technologies for battery management sys-

tems of stationary energy storage systems. Journal of

Power Electronics, 20(6):1526–1540.

Rathod, A. A. and Subramanian, B. (2022). Scrutiny of hy-

brid renewable energy system for control, power man-

agement, optimization and sizing: Challenges and fu-

ture possibilities. Sustainability.

Zhang, J. and Tao, D. (2021). Empowering things with in-

telligence: A survey of the progress, challenges, and

opportunities in artiﬁcial intelligence of things. IEEE

Internet of Things Journal, 8(10):7789–7817.

Zhou, H., Wang, X., and Zhu, R. (2022). Feature selection

based on mutual information with correlation coefﬁ-

cient. Applied Intelligence, 52(5):5457–5474.

SMARTGREENS 2024 - 13th International Conference on Smart Cities and Green ICT Systems