International Journal of Artificial Intelligence and Soft Computing (IJAISC) Vol.1, No.4, September 2023.

SHORT-TERM LOAD FORECASTING USING A

PARTICLE SWARM OPTIMISED MULTI-HEAD
ATTENTION-AUGMENTED CNN-LSTM MODEL

Paapa Kwesi Quansah1 and Edwin Kwesi Ansah Tenkorang2

Department of Electrical/Electronic Engineering, Kwame Nkrumah University of

Science and Technology, Kumasi, Ghana

ABSTRACT
Short-term load forecasting is of utmost importance in the efficient operation and planning of power
systems, given their inherent non-linear and dynamic nature. Recent strides in deep learning have
shown promise in addressing this challenge. However, these methods often grapple with
hyperparameter sensitivity, opaqueness in interpretability, and high computational overhead for real-
time deployment. This paper proposes an innovative approach that effectively overcomes the
aforementioned problems. The approach utilizes the Particle Swarm Optimization algorithm to
autonomously tune hyperparameters, a Multi-Head Attention mechanism to discern the salient features
crucial for accurate forecasting, and a streamlined framework for computational efficiency. The method
was subjected to rigorous evaluation using a genuine electricity demand dataset. The results underscore
its superiority in terms of accuracy, robustness, and computational efficiency. Notably, its Mean
Absolute Percentage Error of 1.9376 marks a significant improvement over existing state-of-the-art
approaches, heralding a new era in short-term load forecasting.

KEYWORDS
Short-Term Load Forecasting, Deep Learning, Particle-Swarm Optimization, Multi-Head Attention,
CNN-LSTM Network, Electricity Demand

1. INTRODUCTION
In contemporary society, electrical energy has emerged as a pivotal resource propelling the
economic and societal progress of nations worldwide. It is extensively utilised in industries,
including manufacturing, mining, construction, and healthcare, among others. The provision
of consistent and high-quality electrical power is not merely a convenience; rather, it is
imperative to sustain investor confidence in economies and foster further development [1].
With the advent of new technological advancements, electricity demand has surged, creating
an urgent need for more cost-effective and reliable power supply solutions [2].

The current energy infrastructure lacks substantial energy storage capabilities in the
generation, transmission, and distribution systems [3]. This deficiency necessitates a precise
balance between electricity generation and consumption. The maintenance of balance is
contingent upon the utilisation of an accurate load forecasting approach. Adapting electricity
generation to dynamically meet shifting demand patterns is paramount; since failure to do so
puts the stability of the entire power system at risk [4]. Moreover, as the world pivots towards
the increased adoption of renewable energy sources [5], power grids have witnessed a
substantial transformation in their composition and structure. This integration of renewable
energy sources, such as wind and solar power, introduces a degree of unpredictability into
energy generation due to the stochastic nature of these sources [6]. Consequently, ensuring a
stable and secure power system operation becomes an even more complex endeavour,
demanding meticulous power planning and precise load forecasting.

1
International Journal of Artificial Intelligence and Soft Computing (IJAISC) Vol.1, No.4, September 2023.

Electric load forecasting is the practice of predicting electricity demand within a specific
region. This process can be categorised into three distinct groups: short-term, medium-term,
and long-term forecasting, depending on the forecasting horizon. Short-term load forecasting
(STLF), which focuses on predicting electricity demand for upcoming hours, a day, or a few
days, serves as the foundation for effective power system operation and analysis. It facilitates
the optimization of the operating schedules of generating units, including their start and stop
times, and their expected output. The accuracy of STLF is of critical importance, as it directly
influences the efficient utilisation of generating units [7]. The absence of accurate short-term
load forecasting can lead to many operational challenges, including load shedding, partial or
complete outages, and voltage instability. These issues can have detrimental effects on
equipment functionality and pose potential risks to human safety.

Short-term load forecasting methods are pivotal in achieving this precision. These methods
can be broadly classified into two main categories: statistical methods and machine learning
methods [8-9]. Machine learning-based load forecasting methods, such as the autoregressive
integrated moving average model (ARIMA) [10], long short-term memory (LSTM) [11],
generative adversarial network (GAN) [12], and convolutional neural network (CNN) [13],
have gained prominence. These machine learning methods excel at capturing complex
nonlinear data features within load patterns [14]. They leverage the ability to discern
similarities in electricity consumption across diverse power supply areas and customer types,
allowing for more accurate and feasible load forecasting through the consideration of spatial-
temporal coupling correlations.

2.1. Motivation
Based on the existing research, the following three shortcomings need to be addressed to
improve the forecasting effect of the spatial–temporal distribution of the system load: (i) the
lack of flexibility and scalability of traditional statistical methods, (ii) the high computational
complexity of deep learning methods, and (iii) the inability of existing methods to capture the
spatial-temporal correlations in load patterns.

Considering these challenges, this paper proposes a novel short-term load forecasting model
that uses a particle swarm-optimised multi-head attention-augmented CNN-LSTM network.
The proposed model employs a particle swarm optimization (PSO) algorithm to identify the
optimal hyperparameters of the CNN-LSTM network. This enhances the model's resilience to
overfitting and its accuracy. Additionally, the multi-head attention mechanism is used to learn
the importance of different features for the forecasting task. Finally, a hybrid CNN-LSTM
Model is used to help the system capture the spatial-temporal correlations in load patterns,
hence enhancing its forecasting accuracy.

2.2. Contributions
The following are the main contributions of the paper:

1. Feature Extraction: To improve efficiency during feature extraction for STLF, PSO
is employed to optimise model hyperparameters, leading to enhanced efficiency in
extracting significant features with lower computational resources.
2. Attention-Augmented Hybrid Model: Given that power demand is impacted by
short-term fluctuations and long-term trends in data, a hybrid model is used to detect
both temporal and extended dependencies, improving accuracy.
3. Performance evaluation: The effectiveness of PSO-A2C-LNet has been validated
using three real-world electricity demand data sets (from Panama, France, and the
US). Testing results demonstrate that the PSO-A2C-LNet outperforms benchmarks in
terms of forecasting performance.

2
International Journal of Artificial Intelligence and Soft Computing (IJAISC) Vol.1, No.4, September 2023.

2.3. Structure of this paper

The rest of the paper is structured as follows. Section II provides comprehensive explanations
and definitions of key terminology. Section III provides an in-depth exposition of the proposed
framework and a comprehensive explanation of its operation. The findings of our tests on the
framework are presented in Section IV. Section V concludes the paper.

2. FRAMEWORK COMPONENTS
2.1. Definitions of Key Terms
2.1.1 Convolutional Neural Network
A Convolutional Neural Network (CNN) is a deep learning model designed primarily for
image-related tasks, but it can also be applied to other grid-like data, such as audio or time
series data. CNNs are especially effective at capturing spatial dependencies within inputs. [14].

The CNN achieves the localization of spatial dependencies by using the following layers:

1. Convolutional Layer: The core operation in a CNN is the convolution operation.

Convolutional layers use learnable filters or kernels to scan the input data in a localised and
overlapping manner. Each filter detects specific features, like edges, textures, or more complex
patterns.

Mathematically, the 2D convolution operation is defined as follows:

Here,

- Y is the filter (kernel) of size M x N.

- X is the input data of size (W, H).

- (i, j) represents the coordinates of the output feature map.

- (m, n) iterates over the filter dimensions.

By sliding the filter across the input, the convolution operation computes feature maps that
highlight different aspects of the input. This process effectively captures spatial dependencies.

2. Pooling Layer: Pooling layers are predominantly used to downsample the feature maps,
reducing their spatial dimensions. Common pooling operations include max-pooling and
average-pooling. Pooling aids in the invariance of network translation and minimises the
computational overhead.

For max-pooling, the operation is defined as:

3
International Journal of Artificial Intelligence and Soft Computing (IJAISC) Vol.1, No.4, September 2023.

where x is the input feature map, and (p, q) represents the pooling window position. Max-
pooling retains the most significant information within the window.

3. Fully Connected Layer: After multiple convolutional and pooling layers, the spatial
dimensions are reduced, and the network connects to one or more fully connected layers, also
known as dense layers. These layers perform classification or regression tasks by learning
high-level representations.

Recognizing and exploiting spatial dependencies in CNNs is facilitated through several key
mechanisms [15]. CNNs utilise local receptive fields, whereby each neuron is connected to a
small region of the input data. This enables neurons to specialise in detecting specific features
within their receptive fields, hence facilitating the network's ability to record spatial
relationships across multiple scales. Additionally, weight sharing is a fundamental aspect of
CNNs, where the same set of filters is applied consistently across the entire input. This weight
sharing allows the network to learn translation invariant patterns, boosting its capacity to grasp
spatial dependencies. Moreover, CNNs employ a hierarchical representation approach, where
deeper layers in the network combine higher-level features derived from lower-level features.
This hierarchical representation aids the network in comprehending complex spatial
dependencies by gradually constructing abstractions. These mechanisms collectively empower
CNNs to effectively model and exploit spatial dependencies in input data.

2.1.2 Long Short-Term Memory Network

The LSTM network is a type of recurrent neural network (RNN) architecture that is designed
to capture and model sequential data while addressing the vanishing gradient problem that
plagues traditional RNNs. LSTMs are particularly effective at locating and modelling long-
term dependencies in sequential data.

LSTMs consist of multiple interconnected cells, each with its own set of gates and memory
cells [16]. The primary components of an LSTM cell are:

Forget Gate (𝑓 ): Controls what information from the previous cell state (𝐶 ) should be
discarded or kept. It takes the previous cell state and the current input (𝑥 ) as input and
produces a forget gate output.

(3)

Input Gate (𝑖 ): Determines what new information should be added to the cell state. It takes
the previous cell state and the current input and produces an input gate output.

(4)

Candidate Cell State (~𝐶 ): This is a candidate new cell state, computed using the current
input and a tanh activation function.

(5)

Cell State Update (𝐶 ): The cell state is updated by combining the information retained from
the previous cell state and the new candidate cell state .

(6)

4
International Journal of Artificial Intelligence and Soft Computing (IJAISC) Vol.1, No.4, September 2023.

Output Gate (𝑜 ): Determines what part of the cell state should be output as the final
prediction. It takes the current input and the updated cell state and produces an output gate
output.

(7)

Hidden State (ℎ ): The hidden state is the output of the LSTM cell, which is used as the
prediction and is also passed to the next time step. It is calculated by applying the output gate
to the cell state.

(8)

LSTMs address the vanishing gradient issue of traditional RNNs by introducing key
components: the cell state (𝐶 ) and the forget gate (𝑓 ) [17]. The forget gate dynamically
adjusts (𝑓 ) to enable LSTMs to remember or discard information from distant time steps,
facilitating the capture of long-term dependencies. Meanwhile, the cell state (𝐶 ) acts as a
memory buffer, accumulating and passing relevant information across time steps, thus
enabling the model to recognize and exploit long-term patterns within input sequences.

2.1.2 Multi-Head Attentional Mechanism

The Multi-Head Attention mechanism [18] is a key component of Transformer-based models,

such as BERT and GPT, used for various natural language processing tasks. It excels at
capturing extremely long-term dependencies in sequences of data.

Multi-Head Attention extends the idea of the self-attention mechanism [19] by employing
multiple attention heads in parallel. Each attention head focuses on different parts of the input
sequence, enabling the model to capture various types of information and dependencies
simultaneously.

The primary components of Multi-Head Attention are as follows:

Query (Q), Key (K), and Value (V) Projections: For each attention head, we project the input
sequence into three different spaces: query, key, and value. These projections are learned
parameters.

Scaled Dot-Product Attention: Each attention head computes attention scores between the
query (Q) and the keys (K) of the input sequence and then uses these scores to weight the
values (V). The attention scores are computed as a scaled dot product:

(9)

Here, 𝑑 is the dimension of the key vectors.

Concatenation and Linear Transformation: After computing the attention outputs for each
head, we concatenate them and apply a linear transformation to obtain the final multi-head
attention output:

(10)

5
International Journal of Artificial Intelligence and Soft Computing (IJAISC) Vol.1, No.4, September 2023.

Where Concat concatenates the outputs from all attention heads, and 𝑊 is a learned linear
transformation.

3. PSO-A2C-LNET ARCHITECTURE

Figure 1. PSO-A2C-LNet Structure Diagram

PSO-A2C-LNet utilises the various aforementioned components to extract more relevant

features of the power load data to provide better predictions. This algorithm effectively
improves the prediction accuracy of STLF. The model architecture is described below.

The model starts with an input layer designed to accept sequential data. Following the input
layer, a Convolutional layer is used to capture temporal spatial patterns in the data.
Subsequently, a bidirectional LSTM layer is employed to model long-term dependencies both
forward and backward, enabling the capture of historical data through time. The crucial Multi-
Head Attention module operates on the output of the first bidirectional LSTM layer, enabling
the model to focus on the most relevant features and learn their importance. To capture
intricate long-term patterns, a second bidirectional LSTM layer is employed. The final LSTM
layer generates a probabilistic value. The anticipated short-term demand in kilowatts per hour
is predicted by the output layer, which consists of a dense layer with one neuron and a linear
activation function. A few Dropout layers were interspersed among the other model's other
layers to combat overfitting. Layer Normalisation is implemented subsequent to the first
bidirectional layer in order to provide consistent and steady training across different inputs.
The hyperbolic tangent (tanh) activation function was used for all LSTM layers.

6
International Journal of Artificial Intelligence and Soft Computing (IJAISC) Vol.1, No.4, September 2023.

To optimise model performance and convergence during training, PSO was employed to fine-
tune critical hyperparameters. Table I shows the optimised hyperparameters and their
corresponding optimization ranges.

Table I

Optimization Ranges

Hyperparameter Range
Adaptive Learning Rate [0.001, 0.1]
Batch Size [1, 128]
Number of Epochs [100, 5000]
Weight Initialization Techniques (Xavier, He, Random)
Loss Metrics (MSE, Cross-Entropy, MAPE)

The specific implementation process of the proposed algorithm is provided in Algorithm 1.

4. RESULTS AND DISCUSSION

This section comprehensively analyses the STLF results by implementing the above model
and testing it extensively on three datasets; ERCOT, RTE, and the Panama Energy Dataset.
Three regression evaluation metrics are introduced for quantitative analysis of the prediction
results. The performance and fitting degree of the different models are measured by the
following indicators:

7
International Journal of Artificial Intelligence and Soft Computing (IJAISC) Vol.1, No.4, September 2023.

where n: Number of Observations, 𝑌 : Actual values at data point i, 𝑌 : Predicted values at data
point i and 𝑌: Mean of the observed values.

Table II

Performance Metrics for Different Models on Various Datasets

Dataset Model 𝑅 MAPE(%) MAE

Panama Dataset A2C-LNet 0.85 2.8 8.1
PSO-A2C-LNet 0.88 1.9 7.3
Hybrid CNN-LSTM 0.92 3.1 8.7
Vanilla CNN 0.81 3.9 11.2
Vanilla LSTM 0.86 3.4 10.0
ERCOT Dataset A2C-LNet 0.92 3.1 8.7
PSO-A2C-LNet 0.87 2.1 7.5
Hybrid CNN-LSTM 0.89 2.6 9.1
Vanilla CNN 0.95 3.2 7.8
Vanilla LSTM 0.91 4.0 8.5
RTE Dataset A2C-LNet 0.78 2.4 12.3
PSO-A2C-LNet 0.86 2.0 7.5
Hybrid CNN-LSTM 0.89 2.9 7.2
Vanilla CNN 0.79 4.2 12.0
Vanilla LSTM 0.81 3.1 11.2

From the results Table II, the PSO-A2C-LNet model consistently stands out. On the Panama
Energy Dataset, it achieves the highest coefficient of determination (𝑅 ) at 0.88, indicating
strong predictive accuracy, along with the lowest mean absolute percentage error (MAPE) of
1.9% and the smallest mean absolute error (MAE) of 7.3, making it the top-performing model.
In the ERCOT Dataset, PSO-A2C-LNet also delivers competitive results with an (𝑅 ) of 0.87,
a MAPE of 2.1%, and a MAE of 7.5. Similarly, on the RTE Dataset, it outperforms other
models with an (𝑅 ) score of 0.86, a lower MAPE of 2.0%, and a MAE of 7.5. These consistent
results suggest that PSO-A2C-LNet exhibits robust predictive capabilities across diverse
datasets.
While PSO-A2C-LNet excels on all datasets, the other models exhibit varying levels of
performance. These comparative results emphasise the importance of model selection based
on the specific dataset and application, with PSO-A2C-LNet emerging as a robust choice for
diverse predictive tasks.
4.1 Comparison of results with results in literature
Table III shows the results of the A2C-LNet and the PSO-A2C-LNet on the testing dataset
compared to other models in scientific literature.

8
International Journal of Artificial Intelligence and Soft Computing (IJAISC) Vol.1, No.4, September 2023.

Table III

Comparison of MAPE Results with Literature

Model Best MAPE(%) Results

A2C-LNet 2.8097
PSO-A2C-LNet (Our Proposed Model) 1.9376
Integrated CNN and LSTM Network [20] 3.49
LSTM network considering attention mechanism [10] 2.26
ANN-IEAMCGM-R [21] 3.59
TCN-LightGBM [22] 2.64
nonAda-GWN [23] 7.42
Ada-GWN [23] 6.83
Stacked XGB-LGBM-MLP [23] 2.69
GRU-CNN Hybrid Neural Network Model [24] 2.8839

5. CONCLUSIONS
In conclusion, this research paper has introduced a novel neural network architecture for short-
term load forecasting, amalgamating Convolutional Neural Network and Long Short-Term
Memory models, reinforced by a Multi-Head Attention Mechanism. Empirical assessments
confirm its superiority over traditional methods and standalone neural network models, with
demonstrated applicability to real-world datasets.

Future work will focus on optimising the proposed architecture, exploring further
hyperparameter tuning, and investigating additional data preprocessing techniques for
enhanced forecasting. Additionally, integrating robust data privacy measures, such as
federated learning or secure enclaves, into the architecture is essential to address emerging
privacy concerns in load forecasting, ensuring secure and privacy-preserving predictions while
advancing the scalability and adaptability of the framework to diverse forecasting challenges
and datasets.

DECLARATION OF COMPETING INTEREST

The authors declare that there is no conflict of interest regarding the publication of this paper.

ACKNOWLEDGEMENTS
The authors would like to thank Professor Philip Yaw Okyere for guiding the research.

9
International Journal of Artificial Intelligence and Soft Computing (IJAISC) Vol.1, No.4, September 2023.

REFERENCES
[1] D. Frederick and A. E. Selase, “The effect of electric power fluctuations on the profitability
and competitiveness of smes: A study of smes within the accra business district of Ghana,”
Journal of Cryptology, vol. 6, pp. 32–48, 2014.
[2] N. D. Rao and S. Pachauri, “Energy access and living standards: Some observations on
recent trends,” Environmental Research Letters, vol. 12, no. 2, p. 025 011,2017.
[3] T. M. Letcher, 11-storing electrical energy, editor (s): Trevor M. Letcher, managing global
warming, 2019.
[4] P. Jiang, F. Liu, and Y. Song, “A hybrid forecasting model based on date-framework strategy
and improved feature selection technology for short-term load forecasting,” Energy, vol. 119,
pp. 694–709, 2017.
[5] O. Ellabban, H. Abu-Rub, and F. Blaabjerg, “Renewable energy resources: Current status,
future prospects and their enabling technology,” Renewable and sustainable energy reviews,
vol. 39, pp. 748–764, 2014.
[6] D. Ahmed, M. Ebeed, A. Ali, A. S. Alghamdi, and S. Kamel, “Multi-objective energy
management of a micro-grid considering stochastic nature of load and renewable energy
resources,” Electronics, vol. 10, no. 4, p. 403, 2021.
[7] Y. Chen, P. B. Luh, C. Guan, et al., “Short-term load forecasting: Similar day-based wavelet
neural networks,” IEEE Transactions on Power Systems, vol. 25, no. 1, pp. 322–330, 2009.
[8] Y. Hu, B. Qu, J. Wang, et al., “Short-term load forecasting using multimodal evolutionary
algorithm and random vector functional link network based ensemble learning,” Applied
Energy, vol. 285, p. 116 415, 2021.
[9] Y. Kim, H.-g. Son, and S. Kim, “Short term electricity load forecasting for institutional
buildings,” Energy Reports, vol. 5, pp. 1270–1280, 2019.
[10] J. Lin, J. Ma, J. Zhu, and Y. Cui, “Short-term load forecasting based on LSTM networks
considering attention mechanism,” International Journal of Electrical Power And Energy
Systems, vol. 137, p. 107 818, 2022.
[11] N. Huang, Q. He, J. Qi, et al., “Multi Nodes interval electric vehicle day-ahead charging load
forecasting based on joint adversarial generation,” International Journal of Electrical Power
And Energy Systems, vol. 143, p. 108 404, 2022.
[12] T.-Y. Kim and S.-B. Cho, “Predicting residential energy consumption using cnn-lstm neural
networks,”Energy, vol. 182, pp. 72–81, 2019.
[13] Y. Liu, Q. Wang, X. Wang, et al., “Community enhanced graph convolutional networks,”
Pattern Recognition Letters, vol. 138, pp. 462–468, 2020.
[14] Z. Li, F. Liu, W. Yang, S. Peng, and J. Zhou, “A survey of convolutional neural networks:
Analysis, applications, and prospects,” IEEE transactions on neural networks and learning
systems, 2021.
[15] Y. Li, H. Zhang, and Q. Shen, “Spectral–spatial classification of hyperspectral imagery with
3d convolutional neural network,” Remote Sensing, vol. 9, no. 1, p. 67, 2017.
[16] R. C. Staudemeyer and E. R. Morris, “Understanding lstm–a tutorial into long short-term
memory recurrent neural networks,” arXiv preprint arXiv:1909.09586, 2019.
[17] S. Chandar, C. Sankar, E. Vorontsov, S. E. Kahou, and Y. Bengio, “Towards non-saturating
recurrent units for modelling long-term dependencies,” in Proceedings of the AAAI
Conference on Artificial Intelligence, vol. 33, 2019, pp. 3280–3287.
[18] J.-B. Cordonnier, A. Loukas, and M. Jaggi, “Multihead attention: Collaborate instead of
concatenate,”arXiv preprint arXiv:2006.16362, 2020.

10
International Journal of Artificial Intelligence and Soft Computing (IJAISC) Vol.1, No.4, September 2023.

[19] P. Ramachandran, N. Parmar, A. Vaswani, I. Bello, A. Levskaya, and J. Shlens, “Stand-alone

self-attention in vision models,” Advances in neural information processing systems, vol. 32,
2019.
[20] S. H. Rafi, Nahid-Al-Masood, S. R. Deeba, and E. Hossain, “A short-term load forecasting
method using integrated cnn and LSTM network,” IEEE Access, vol. 9, pp. 32 436–32 448,
2021. DOI: 10 . 1109 / ACCESS .2021.3060654.
[21] P. Singh, P. Dwivedi, and V. Kant, “A hybrid method based on neural network and improved
environmental adaptation method using controlled gaussian mutation with real parameters for
short-term load forecasting,”Energy, vol. 174, pp. 460–477, 2019, ISSN: 0360-5442. DOI:
https://doi.org/10.1016/j.energy.2019.02.141. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S0360544219303408.
[22] Y. Wang, J. Chen, X. Chen, et al., “Short-term load forecasting for industrial customers
based on TCN LightGBM,” IEEE Transactions on Power Systems, vol. 36, no. 3, pp. 1984–
1997, 2021. DOI: 10.1109/TPWRS.2020.3028133.
[23] W. Lin, D. Wu, and B. Boulet, “Spatial-temporal residential short-term load forecasting via
graph neural networks,” IEEE Transactions on Smart Grid, vol. 12, no. 6, pp. 5373–5384,
2021. DOI: 10.1109/TSG.2021.3093515.
[24 L. Wu, C. Kong, X. Hao, and W. Chen, “A Short-Term load forecasting method based on
GRU-CNN hybrid neural network model,” Mathematical Problems in Engineering, vol.
2020, p. 1 428 104, Mar. 2020.

Authors
Paapa Kwesi Quansah is a Machine Learning
Researcher and a Teaching and Research
Assistant of Electrical/Electronics Engineering
Photo
at the Kwame Nkrumah University of Science
and Technology. He is also the founder of Rune,
Inc. a start-up focused on low resource machine
learning systems capable of providing
personalised recommendations to farmers in the
Sub-Saharan region of Africa.His research
primarily focuses on learning mechanisms that
allow autonomous agents to behave intelligently
and coordinate actions for the advancement and
use of real-world autonomous systems.
Edwin Kwesi Ansah Tenkorang completed his
Photo
Bachelor's degree in Electrical/Electronic
Engineering in 2022 from the Kwame Nkrumah
University of Science and Technology
(K.N.U.S.T). He is an aspiring research who has
worked on various research projects spanning
applied machine learning, wireless network
efficiency, power system anomaly detection and
networking paradigms. Edwin is currently a
Cybersecurity Research Assistant at the
University Information Technology Services
centre, K.N.U.S.T, where he explores applicable
open-source security solutions.

11