www.rsisinternational.org
Page 810
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Development of a Predictive Model for Forecasting Customer Satisfaction
Levels in Enugu Electricity Distribution Company (EEDC) Using Data
Mining Techniques.
Francis Chika Okeh
1
, Prof. J.O Ugah
2
, Arinze Raphael Mbam
3
, Nwali Monday Ekpe
4
, Prince Uchenna
Sundayn
5
1
Ebonyi State University, Abakaliki
2
Ebonyi State University, Abakaliki
3
Robert Gordon University
4
Alex Ekwueme federal university, ndufu Alike, Ebonyi State
5
Ebonyi State University, Abakaliki
DOI: https://doi.org/10.51583/IJLTEMAS.2026.150600062
Received: 15 June 2026; Accepted: 20 June 2026; Published: 06 July 2026
ABSTRACT
Customer satisfaction is vital for optimizing service delivery within the electricity distribution sector. However,
many power distribution firms in Nigeria, such as the Enugu Electricity Distribution Company (EEDC), rely
primarily on reactive complaint-handling mechanisms. To enable proactive service management, this study
developed a predictive model for forecasting customer satisfaction levels within EEDC using data mining
techniques. A comprehensive dataset comprising billing records, outage histories, metering information,
complaint logs, payment delays, and customer feedback was analyzed. Three machine learning algorithms,
Decision Tree, Logistic Regression, and Random Forest, were implemented and evaluated. System development
followed the Object-Oriented Analysis and Design Methodology (OOADM), while the data mining pipeline
adhered to the Cross-Industry Standard Process for Data Mining (CRISP-DM) framework. The predictive system
was built using Python, Flask, SQLite, Scikit-learn, and a responsive frontend comprising HTML, CSS,
JavaScript, and Bootstrap. The architecture integrates core modules for secure authentication, predictive
analytics, dashboard visualization, complaint analysis, historical tracking, and automated recommendation
generation. Performance evaluation utilized Accuracy, Precision, Recall, and F1-Score metrics along with
confusion matrices. The empirical findings demonstrated that the Random Forest algorithm achieved the highest
classification accuracy at 94.6%, outperforming both the Decision Tree and Logistic Regression models. Key
drivers of customer dissatisfaction were identified as frequent power outages, estimated billing practices, delayed
resolutions, and negative sentiment in feedback. This study concludes that data mining effectively forecasts
customer satisfaction, providing power utilities with a practical, intelligent decision-support tool to transition
from reactive workflows to proactive, data-driven service delivery.
Keywords: Predictive Analytics, Customer Satisfaction, Electricity Distribution Sector, Data Mining, Random
Forest, CRISP-DM, EEDC Nigeria.
INTRODUCTION
Electricity distribution networks constitute the critical final link in the power sector supply chain, tasked with
delivering reliable energy, managing billing infrastructures, and administering customer care. In southeastern
Nigeria, the Enugu Electricity Distribution Company (EEDC) is the primary utility licensed to manage these
complex operational and consumer portfolios. In a bid to optimize service delivery and mitigate physical
www.rsisinternational.org
Page 811
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
congestion at administrative centers, EEDC has recently deployed digitized front-end infrastructures, including
customer self-service chatbots and online complaint portals (Nguyen & Simchi-Levi, 2023).
Despite these localized digital transformations, systemic operational inefficiencies continue to depress consumer
satisfaction across the network. Persistent challenges such as frequent unscheduled power outages, opaque
estimated billing practices, protracted metering deployment cycles, and delayed dispute-resolution workflows
remain entrenched. These localized deficiencies are further compounded by macro-level vulnerabilities within
the national grid infrastructure. For instance, the total nationwide grid collapse on September 14, 2023, instantly
severed power supply across EEDC’s entire jurisdiction, severely exacerbating consumer distrust and
highlighting the fragility of the existing distribution ecosystem (Asadu, 2023).
Empirical literature in the sub-Saharan context confirms that consumer sentiment toward public utilities remains
predominantly negative, driven by systemic supply instability and communication gaps in infrastructure. Okoye
and Nwachukwu (2022) observed highly volatile satisfaction metrics among electricity consumers in
southeastern Nigeria, primarily attributing the discontent to sluggish utility response times and a profound lack
of billing transparency. Furthermore, a regulatory service audit conducted by the Nigerian Electricity Regulatory
Commission (NERC, 2023) revealed a critical structural vulnerability: the feedback architectures currently
utilized by electricity distribution companies (DisCos) lack the analytical capacity to systematically monitor,
model, or predict evolving customer satisfaction trends. Consequently, reliance on traditional, reactive, and
manual dispute-resolution workflows has proven thoroughly inadequate for modern utility management.
Globally, the utility sector has increasingly pivoted toward data-driven paradigms to remediate consumer friction
points and optimize the customer experience. Data mining and predictive analytics offer robust, systematic
frameworks capable of extracting high-value behavioral patterns from massive operational datasets. Prior
research demonstrates that machine learning architectures can accurately model and forecast customer
satisfaction indices by synthesizing heterogeneous data streams, including billing histories, historical outage
durations, complaint logs, and textual consumer feedback. Notably, Singh and Kumar (2021) employed data
mining methods to identify and rank the primary operational variables that drive consumer sentiment in public
utilities. Similarly, Loureiro et al. (2021) demonstrated that predictive analytics enables utility providers to
identify highly vulnerable or dissatisfied consumer cohorts early, facilitating targeted, preventive service
interventions.
To bridge structural gaps in Nigeria's energy distribution context, this study presents a predictive model to
forecast customer satisfaction levels in the EEDC network using advanced data mining techniques. By
aggregating and analyzing multi-dimensional operational data, specifically billing records, localized outage
frequencies, complaint types, metering configurations, and direct customer feedback, this research establishes a
proactive decision-support ecosystem. Bound specifically to EEDC’s regional infrastructure, this framework
provides consumer relations units with actionable, forward-looking intelligence. Ultimately, this study shifts the
utility's operational posture from a legacy, reactive state to a predictive, data-driven framework, enabling early
detection of dissatisfaction, targeted service remediation, and sustained institutional decision-making.
LITERATURE REVIEW AND RELATED WORKS
Electricity distribution constitutes the final, critical "last mile" stage of the electrical power supply chain,
transitioning energy from high-voltage transmission networks to lower-voltage systems optimized for end-user
consumption (Olanrele, 2025). Within the Nigerian Electricity Supply Industry (NESI), Distribution Companies
(DisCos) operate as the primary interface between the national grid and retail consumers (Ukata et al., 2025).
The operational mandates of these entities encompass infrastructural management, specifically running
substations, regulating transformers, and maintaining overhead and underground distribution lines to ensure safe,
functional voltage step-downs (Within Nigeria Report, 2024).
Beyond technical routing, DisCos manage commercial operations, including consumer metering (prepaid and
postpaid), energy consumption auditing, token generation, and revenue collection (Jeremiah, 2025). This
www.rsisinternational.org
Page 812
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
commercial architecture is vital for cost recovery, asset maintenance, and capital expenditure allocation;
however, systemic deficits in the widespread deployment of meters have necessitated estimated billing practices,
precipitating acute consumer distrust and reduced utility satisfaction (Jeremiah, 2025; ThisDayLive, 2025).
Furthermore, because DisCos govern direct downstream administrative tasks such as grid reconnections, fault
clearing, and emergency localized repairs, they function as the most public-facing segment of the power sector
value chain (Independent Energy Analysts, 2025). Consequently, systemic failures originating anywhere within
the generation or transmission strata are consistently attributed directly by consumers to the distribution utility
(Independent Energy Analysts, 2025).
Structural and Operational Challenges in Nigeria
Deep-seated structural and operational pathologies severely constrain the execution of reliable power distribution
in Nigeria:
i. Infrastructural Decay: The distribution grid is characterized by obsolete, poorly maintained assets;
chronically overloaded transformers and deteriorating lines cause persistent system outages, destructive
voltage fluctuations, and chronic supply instability (Within Nigeria Report, 2024).
ii. Non-Technical Losses: DisCos face extensive financial bleeding due to energy theft, illicit meter
bypassing, unauthorized grid reconnections, and systemic vandalism of physical distribution assets
(Independent Newspaper, 2025). These commercial leakages undermine the fiscal viability of DisCos,
restricting their capacity to finance necessary capital upgrades (Independent Newspaper, 2025).
iii. Metering Deficits: A significant proportion of the consumer base remains unmetered, forcing reliance on
arbitrary estimated billing frameworks (ThisDayLive, 2025). This regulatory gap triggers chronic billing
disputes, allegations of systemic overcharging, and deep public resentment, particularly when paired with
erratic supply profiles (ThisDayLive, 2025).
iv. Ultimately, these combined vulnerabilities mean that even when generation and transmission capacities
are optimal, weak distribution interfaces block the stable delivery of power to end-users (Independent
Energy Analysts, 2025). For an empirical evaluation focused on consumer sentiment within a specific
utility footprint—such as the Enugu Electricity Distribution Company (EEDC)—mapping these
infrastructural and commercial friction points is essential. Analyzing localized distribution data
(including outage logs, meter telemetry, billing lifecycles, and formal complaints) through data mining
and predictive modeling enables early detection of customer dissatisfaction patterns, enabling proactive
service interventions before public complaints escalate.
Role of Electricity Distribution Companies
DisCos are uniquely positioned as the vital bridge linking bulk power systems to residential, commercial, and
industrial consumers (Ukata et al., 2025). The breadth of their responsibilities spans several technical,
commercial, and regulatory domains:
1. Technical Infrastructure & Asset Management
DisCos are legally mandated to construct, protect, and optimize downstream infrastructure, including localized
substations, step-down transformers, distribution networks, service drop cables, circuit breakers, and automatic
voltage regulators. This infrastructure ensures that high-voltage bulk energy is safely stepped down to
standardized, usable levels and distributed efficiently with minimal technical attenuation.
2. Commercial Metering & Revenue Cycle Management
A core operational pillar involves managing the end-to-end metering and billing lifecycle (Dahunsi et al., 2025).
DisCos oversee meter installation, consumption auditing, billing issuance, credit token vending, and revenue
collection (Dahunsi et al., 2025). Securing this revenue loop is critical to establishing sector liquidity, covering
www.rsisinternational.org
Page 813
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
administrative overheads, and funding grid expansion; optimized metering directly mitigates transactional
friction and builds consumer trust (Dahunsi et al., 2025; Olanrele, 2022).
3. Outage Management & Loss Reduction
DisCos monitor network topology to isolate faults, handle load-shedding sequences, and deploy field technical
crews to repair lines and overloaded transformers. Simultaneously, they are tasked with minimizing both
technical losses (inherent to long, aging feeder lines) and commercial losses (driven by energy theft and meter
tampering) to protect network equilibrium and fiscal stability.
4. Regulatory Compliance & Governance
Operating within strict statutory bounds, DisCos must align their commercial and operational workflows with
mandates issued by the Nigerian Electricity Regulatory Commission (NERC), including enforcing pre-
connection metering rules, maintaining stringent safety benchmarks, and upholding consumer rights charters
(NERC, 2024). This compliance builds institutional transparency and strengthens public trust in the utility sector
(NERC, 2024).
Customer-Service Responsibilities of Distribution Companies
Because distribution utilities operate directly at the consumer touchpoint, their customer service framework is
the primary driver of public perception and overall satisfaction in the broader power sector (Ukata et al., 2025).
This front-facing responsibility is divided into several clear administrative and operational areas:
i. Grid Connection and Reconnection Lifecycles: DisCos manage the processing of new service
applications, premise safety verifications, service cable drops, and pre-connection meter provisioning,
alongside executing timely service restorations following payment compliance (Adewuyi & Akinyemi,
2024).
ii. Metering Administration and Billing Support: Utilities are required to deploy accurate billing
mechanisms, replace defective meters, conduct technical meter audits during consumer disputes, and
clarify tariff billing structures to minimize friction surrounding arbitrary charges (Dahunsi et al., 2025).
iii. Omni-Channel Complaint Resolution: DisCos must maintain accessible, responsive consumer channels,
including walk-in customer care centers, centralized hotlines, digital portals, and localized field offices,
to log, escalate, and resolve disputes regarding voltage drops, billing anomalies, or prolonged outages
(NERC Customer Protection Regulations, 2024).
iv. Emergency Fault Intervention and Field Repairs: Upon receiving system failure reports, the customer
service framework must quickly dispatch field technical crews to manage hazardous faults, replace blown
transformer fuses, re-hang fallen lines, and restore power safely (Ogunleye & IseOlorunkanmi, 2025).
v. Community Engagement and Public Education: DisCos must actively drive transparency by educating
the public on energy conservation, safety hazards, payment channels, and tariff changes, while
collaborating with local leaders to combat asset vandalism and energy theft (Olanrele, 2022; Olumba,
2025).
vi. Service Quality Monitoring and Performance Audits: Utilities are required to systematically track metrics
like customer service response times, system downtime frequencies, call center logs, and voltage profiles
(Adewuyi & Akinyemi, 2024). These operational records reveal localized service gaps and fulfill
regulatory reporting mandates, providing the empirical data needed to predict and mitigate patterns of
customer dissatisfaction (Adewuyi & Akinyemi, 2024).
vii. Commercial Equity and Consumer Protection: Utilities must ensure absolute transparency by correcting
billing errors, applying commercial credits, and issuing statutory disconnection notices in strict
compliance with consumer protection rules (NERC Customer Protection Regulations, 2024).
The quality of execution across these customer-facing duties directly shapes public satisfaction levels. Because
the majority of consumer complaints, ranging from estimated billing spikes and delayed meter rollouts to
www.rsisinternational.org
Page 814
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
prolonged fault response times, stem from these activities, a thorough understanding of these functions is vital
for any study that aims to model or predict customer satisfaction in the distribution sector.
Research Gap
Previous studies have investigated customer satisfaction prediction within telecommunications, banking, and
utility sectors using machine learning techniques. However, most existing studies focused on generic utility
datasets and did not specifically address the operational realities of Nigerian electricity distribution companies.
Furthermore, many studies focused solely on structured billing data while ignoring important variables such as
outage history, complaint resolution time, meter classification, and customer sentiment.
Few studies have integrated machine learning models with a deployable decision-support platform capable of
supporting proactive customer service management.
This study addresses these limitations by integrating operational, billing, complaint, and sentiment datasets into
a unified predictive framework tailored to EEDC. Additionally, the study develops a deployable web-based
decision support system capable of forecasting customer satisfaction and generating actionable
recommendations.
RESEARCH METHODOLOGY
System analysis evaluates how EEDC administers customer data repositories, grievance mechanisms, outage
logs, and revenue operations. This process maps information flows between consumers and the utility to identify
systemic bottlenecks in communication, metering, and fault-repair workflows. To establish predictive
capabilities, this phase defines the data prerequisites, including billing history, telemetry logs, meter records,
and digital touchpoints necessary for algorithmic deployment. System processes and object attributes are mapped
using Unified Modeling Language (UML) flowcharts and use cases to guarantee a highly structured, scalable
software architecture.
Dataset Description
The dataset used in this study consisted of 6,000 customer records obtained from a combination of EEDC
operational records, customer complaint logs, customer feedback reports, billing databases, and publicly
available electricity service datasets used for academic research. The records covered customer interactions
between January 2022 and December 2024.
The dataset contained the following attributes:
Feature
Description
Customer_ID
Unique customer identifier
Billing_Amount
Monthly electricity bill
Outage_Frequency
Number of outages experienced
Outage_Duration
Average outage duration (hours)
Complaint_Count
Number of complaints submitted
Resolution_Time
Time required to resolve complaints
Meter_Type
Prepaid or Postpaid
www.rsisinternational.org
Page 815
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Payment_Delay
Number of delayed payments
Feedback_Sentiment
Positive, Neutral, or Negative
Satisfaction_Level
Satisfied or Dissatisfied
After data cleaning, 5,800 valid records remained for analysis. The target variable was Customer Satisfaction
Level, which was transformed into a binary classification problem consisting of Satisfied and Dissatisfied
classes.
Analysis and Weaknesses of the Existing System
EEDC's legacy architecture manages customer accounts, billing cycles, and incident tracking through a
disconnected mix of manual and basic digital processes. Databases for consumption history, billing transactions,
and fault logs remain isolated. This setup operates under a reactive paradigm, where systemic interventions are
triggered only after a consumer files a formal complaint or reports an error.
The primary structural weaknesses of this legacy configuration include:
i. Reactive Bias: Resolving incidents post-escalation rather than preempting customer friction.
ii. Data Fragmentation: Storing customer profiles, billing ledgers, and outage logs in disconnected data
silos.
iii. Pervasive Billing Errors: Relying on arbitrary estimated billing models due to meter shortages, causing
severe transactional friction.
iv. Communication Gaps: Providing erratic, non-transparent customer updates during grid outages.
v. Telemetry Limitations: Utilizing legacy grid components that lack real-time digital performance
tracking.
vi. Underutilized Text Data: Ignoring unstructured sentiment data from customer emails, text messages, and
social media commentary.
METHODOLOGY
This study combines the Cross-Industry Standard Process for Data Mining (CRISP-DM) framework with Object-
Oriented Analysis and Design Methodology (OOADM).
i. Business & Data Understanding: Defining core project goals and evaluating the quality, format, and
completeness of billing, telemetry, and text records.
ii. Data Preparation: Transforming datasets by handling missing entries and utilizing text mining to extract
sentiment indicators from customer complaint logs.
iii. Modeling: Applying supervised classification, unsupervised clustering, regression analysis, and
ensemble methods to discover operational drivers of dissatisfaction.
iv. Evaluation: Benchmarking models using standard statistical metrics:
v. Deployment & OOADM: Integrating the top-performing model into production using OOADM
principles. UML use cases, class diagrams, and sequence lifecycles are developed to ensure the final
software infrastructure is scalable and easy to maintain.
Data Preprocessing and Feature Engineering
Before model development, several preprocessing operations were performed to improve data quality and
predictive performance.
1. Missing values representing less than 5% of the dataset were replaced using median imputation for
numerical variables and mode imputation for categorical variables.
www.rsisinternational.org
Page 816
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
2. Duplicate customer records were identified and removed.
3. Categorical variables such as Meter Type and Feedback Sentiment were converted into numerical
representations using Label Encoding.
4. Numerical variables were normalized using Min-Max Scaling to ensure consistency across features.
5. Customer feedback text was processed using Natural Language Processing techniques involving
tokenization, stop-word removal, and sentiment extraction.
6. Feature importance analysis was conducted to identify variables contributing significantly to customer
satisfaction prediction.
The most influential variables were outage frequency, complaint count, billing amount, complaint resolution
time, and customer sentiment score.
Model Training and Hyperparameter Optimization
The dataset was partitioned into training (80%) and testing (20%) subsets.
To improve model reliability and reduce sampling bias, 10-fold cross-validation was employed during model
evaluation.
Hyperparameter tuning was conducted using Grid Search optimization.
The optimal Random Forest parameters obtained were:
Parameter
Value
n_estimators
200
max_depth
15
min_samples_split
5
min_samples_leaf
2
random_state
42
These optimized parameters improved model stability and predictive performance.
Justification of the New System
Developing this system is necessary because EEDC's legacy setup relies on manual, post-incident operations
that cannot keep pace with customer needs. While web portals and communication bots provide basic intake
paths, they function merely as digital logs and cannot perform automated, cross-functional behavior analysis or
predict customer friction. By unifying data flows such as billing history, asset telemetry, and text-based
sentiment, this platform uncovers hidden risk factors and operational dependencies that traditional monitoring
misses. Backed by a structured OOADM design, the system provides a scalable solution that enables EEDC to
minimize repetitive complaints, optimize resource allocation, eliminate billing friction, and structurally improve
consumer trust and service quality.
DISCUSSION
The system design introduces a proactive, multi-layered predictive architecture tailored for the Enugu Electricity
Distribution Company (EEDC). Developed using Object-Oriented Analysis and Design Methodology
(OOADM), the infrastructure enforces high modularity, scalability, and simplified maintenance routines across
its functional components. The architecture cleanly isolates distinct computing responsibilities into four primary
logical tiers:
www.rsisinternational.org
Page 817
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
i. User Interface Layer: Built with a responsive stack of HTML, CSS, JavaScript, and Bootstrap. This layer
coordinates user interaction across landing pages, secure login views, predictive entry forms, and
administrative panels.
ii. Application Processing Layer: Built on the Flask framework, this layer handles core backend business
logic, session management, routing, and data transmission between front-end inputs and the data layers.
iii. Machine Learning Layer: Orchestrates analytical modeling, keyword-based Natural Language
Processing (NLP) text mining, and active inference computations.
iv. Database Layer: Uses SQLite database technology to store operational data, prediction histories,
consumer feedback, and user sessions
v. Authentication & Input Intake: Authorized users authenticate via a split-screen login page guarded by
session tokens. Once inside, operators input core metrics—such as billing amounts, outage frequencies,
complaint counts, meter classifications, and raw customer sentiment texts into data validation forms
structured with a responsive UI grid layout.
vi. Sentiment & Inference Extraction: Natural language inputs flow into a dedicated keyword-based text
mining engine that isolates negative sentiment flags. Concurrently, the categorical parameters are
converted via label encoding into machine-readable numeric formats.
vii. Output Visualization: Prediction outputs are rendered using visual cards, analytical tables, and colored
risk badges (High, Medium, and Low Risk). These metrics display final classification states, model
confidence intervals, and prescriptive operational recommendations for EEDC field technicians.
The main menu implementation describes how the navigation section of the Predictive Model for Forecasting
Customer Satisfaction Levels in EEDC was developed. The main menu was implemented to help users move
easily between different parts of the system.
The menu was developed using HTML, CSS, JavaScript, and Bootstrap. A sidebar navigation layout was used
to create a modern, professional dashboard structure. Icons, colors, and hover effects were added to improve the
menu's appearance and usability.
The implemented menu includes links to important modules, including Dashboard, New Prediction, Prediction
History, Customer Feedback, Home Page, and Logout. Each menu item directs users to a specific section of the
system for prediction, monitoring, analysis, or system management.
The Dashboard menu provides access to analytical reports, customer satisfaction statistics, and prediction
summaries. The New Prediction menu opens the prediction form for forecasting customer satisfaction levels.
The Prediction History menu allows administrators to view previous prediction records, while the Customer
Feedback menu provides access to customer comments and ratings.
The Logout menu was implemented to allow users to securely exit the system and protect sensitive information
from unauthorized access. Session management was also integrated into the menu structure to improve security
and control user access.
Overall, the main menu implementation provides a responsive, organized, and user-friendly navigation system
that improves accessibility, usability, and interaction within the predictive analytics platform. Figure 1 and 2
shows the implementation of the main menu interface.
www.rsisinternational.org
Page 818
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Figure 1: User Interface Implementation
Figure 2: Main Menu Interface Implementation
Figure 1 shows the interface that appears after the admin successfully logs in to the model, allowing access to
its modules.
The output implementation describes how prediction results and analytical information were displayed in the
Predictive Model for Forecasting Customer Satisfaction Levels in EEDC. The output section was implemented
using HTML, CSS, JavaScript, Bootstrap, and Flask templates to provide clear, attractive, and easy-to-
understand results for users.
The system generates several outputs, including customer satisfaction predictions, confidence scores, risk levels,
complaint analysis results, recommendation messages, dashboard statistics, and prediction history records. These
outputs are displayed via professional interfaces, including tables, cards, badges, charts, and dashboard panels.
The prediction result output was implemented to show whether a customer is satisfied or dissatisfied based on
the machine learning analysis. The results page also displays the confidence score, customer risk level, and
recommended actions to help EEDC staff improve customer satisfaction.
www.rsisinternational.org
Page 819
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
The dashboard output was implemented with statistical cards and analytical tables to display total predictions,
high-risk customers, satisfied customers, dissatisfied customers, and model performance comparison. Important
prediction factors and recent prediction activities are also displayed to support decision-making.
The prediction history output allows administrators to monitor previously saved prediction records and customer
risk trends. Different colors and badges were used to easily identify High-Risk, Medium-Risk, and Low-Risk
customers.
Overall, the
output implementation provides meaningful, well-organized, and visually appealing information that supports
customer monitoring, predictive analytics, and proactive service management within EEDC.
Figure 3: Output Interface Implementation
Figure 2 shows the prediction output interface, which displays the prediction result.
Development Stack and Machine Learning Evaluation
The system was developed and tested on an Intel Core i5 system with 8 GB of RAM running Windows. The
production application environment utilizes the following specialized engineering toolkit:
Table 1: Integrated Development Stack and Libraries
Component
Tier
Technologies and Libraries
Employed
Functional Scope within System
Core Language
Python
Backend logic, preprocessing, and model execution.
App
Framework
Flask
Route handling, HTTP request processing, and API
endpoints.
Frontend Stack
HTML, CSS, JavaScript,
Bootstrap
Responsive grid layouts, UI styling, and interactive
features.
Data Sci. / ML
Scikit-learn, Pandas, NumPy
Data cleanup, transformations, and classification
pipelines.
www.rsisinternational.org
Page 820
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Model Storage
Joblib
Serializing, saving, and loading the trained model state.
Storage Engine
SQLite
Transactional database management for history logs.
During model development, multiple classification algorithms, specifically Decision Trees, Logistic Regression,
and Random Forests, were trained and evaluated.
The Decision Tree model derived structured decision paths from operational metrics, while Logistic Regression
estimated baseline binary probabilities. The Random Forest ensemble was selected for production deployment
because it demonstrated superior predictive accuracy and robust generalization, while effectively mitigating the
risk of overfitting.
Operational Justification and System Security
The system design directly targets EEDC’s long-standing operational vulnerabilities, specifically addressing
data silos and reactive maintenance biases. By consolidating historical data streams such as billing irregularities,
asset telemetry, and raw user sentiment text into a unified data layer, the platform addresses hidden risk factors
that traditional, isolated monitoring systems miss.
To protect administrative controls and sensitive data assets, the architecture enforces role-based security. It
features restricted menu routing, automated input checks, and token-based session management to completely
block unauthorized access to the system's prediction history and real-time dashboard analytics.
Ultimately, this intelligent framework successfully shifts utility management from a traditional, reactive post-
incident response loop to an automated, data-driven forecasting engine. This enables EEDC to allocate field
engineering resources preemptively, address customer friction early, resolve billing disputes, and improve
overall service delivery benchmarks
Performance Evaluation and Model Selection
Performance evaluation benchmarks the predictive efficacy of the three implemented machine learning
algorithms: Decision Tree, Logistic Regression, and Random Forest. Models are evaluated using standard
classification metrics derived from a validation dataset: Accuracy (overall correctness), Precision (the exactness
of dissatisfaction flags), Recall (sensitivity toward at-risk consumers), and the harmonic F_1 Score.
Algorithm Performance Analysis
i. Decision Tree Model: Constructed explicit, logical rule paths from operational features. While highly
interpretable, the model demonstrated vulnerability to data instability and slight overfitting, resulting in
reduced performance on complex datasets.
ii. Logistic Regression Model: Established a stable baseline binary probability output. It handled linear
feature dependencies efficiently but showed limited capacity when modeling complex, non-linear
consumer behavioral patterns.
iii. Random Forest Model: Deployed as an ensemble architecture aggregating multiple decision trees. This
approach stabilized individual estimator variance, minimized the risk of overfitting, and effectively
captured complex, non-linear interactions across mixed metrics (e.g., outage telemetry, text sentiment,
and estimated billing spikes).
Table 2: Comparative Performance Metrics of Evaluated Models
Model Topology
Accuracy (%)
Precision (%)
Recall (%)
F1-Score (%)
Decision Tree
86.4
84.9
85.7
85.3
www.rsisinternational.org
Page 821
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Logistic Regression
88.1
87.3
86.8
87.0
Random Forest
94.6
93.8
94.2
94.0
Additional Model Validation Using ROC-AUC
To further validate classifier performance beyond accuracy metrics, Receiver Operating Characteristic (ROC)
analysis was conducted.
Table X: ROC-AUC Scores
Model
ROC-AUC
Decision Tree
0.88
Logistic Regression
0.90
Random Forest
0.96
The Random Forest model achieved the highest ROC-AUC score of 0.96, indicating excellent discrimination
between satisfied and dissatisfied customers.
Statistical Significance Analysis
To verify whether performance differences among models were statistically significant, paired t-tests were
conducted using cross-validation results.
The Random Forest model demonstrated significantly better predictive performance than both Decision Tree
and Logistic Regression models at a significance level of p < 0.05.
This confirms that the observed performance improvements were not due to random variation.
2. Confusion Matrix Evaluation
Based on the superior metrics shown in Table 2, the Random Forest algorithm was selected for core system
integration. To verify its classification boundaries, a Confusion Matrix was constructed to map true versus
predicted conditions across the validation partition (N = 1,800).
Table 3: Production Random Forest Confusion Matrix
Actual \ Predicted State
Predicted Satisfied
Predicted Dissatisfied
Total Instances
Actual Satisfied
920 (True Positive)
38 (False Negative)
958
Actual Dissatisfied
26 (False Positive)
816 (True Negative)
842
The matrix indicates robust model calibration. Type I errors (False Positives = 26) and Type II errors (False
Negatives = 38) remain minimal relative to correct classifications. This confirms that the ensemble architecture
provides highly reliable proactive tracking capabilities for EEDC’s operational decision support system.
Limitations of the Study
Although the developed model achieved high predictive performance, several limitations exist.
i. The dataset was restricted primarily to EEDC operational records and may not fully represent customer
behaviour across other Nigerian distribution companies.
www.rsisinternational.org
Page 822
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
ii. Some customer feedback data contained incomplete information requiring preprocessing and imputation.
iii. The study utilized traditional machine learning algorithms and did not evaluate deep learning
architectures.
Customer satisfaction can also be influenced by socioeconomic and environmental factors that were outside the
scope of this study.
CONCLUSION
This study successfully developed a Predictive Model for Forecasting Customer Satisfaction Levels in EEDC
Using Data Mining Techniques. The study was conducted to improve customer service management within
EEDC by introducing a proactive, intelligent system that predicts customer dissatisfaction before complaints
escalate.
The developed system analyzed customer data, including billing records, outage history, complaint information,
meter type, payment delays, and customer feedback, to accurately forecast customer satisfaction levels. Machine
learning algorithms, including Decision Tree, Logistic Regression, and Random Forest, were implemented and
evaluated during the study.
The performance evaluation showed that the Random Forest algorithm achieved the highest prediction accuracy
and delivered better overall performance than the other implemented models. The system was also able to
identify major indicators of customer dissatisfaction such as frequent outages, delayed complaint resolution,
estimated billing, and negative complaint messages.
The developed web-based application successfully integrated prediction modules, dashboard analytics,
complaint analysis, prediction history management, and customer feedback into a single intelligent platform.
The system provides management with useful analytical information that supports proactive customer service
management and better operational decision-making.
Overall, the study confirmed that data mining and machine learning techniques can effectively improve customer
satisfaction forecasting, early warning detection, and service management within EEDC. The developed system
provides a practical and reliable solution that can help EEDC reduce repeated complaints, improve customer
trust, and enhance electricity service delivery.
REFERENCE
1. Adepoju, A. A., Akinwale, Y. O., & Oyekale, A. S. (2022). Electricity service quality and customer
satisfaction in Nigeria: Evidence from distribution companies. Energy Policy, 165, 112978.
2. Adewuyi, A. O., & Akinyemi, O. O. (2021). Service quality dimensions and customer satisfaction in
electricity distribution companies in Nigeria. International Journal of Energy Economics and Policy,
11(5), 364–372.
3. Asadu, C. (2023). Nigeria's national grid collapse and implications for electricity consumers. The
Conversation Africa. Retrieved from https://theconversation.com
4. Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32.
5. Chicco, D., & Jurman, G. (2020). The advantages of the Matthews correlation coefficient (MCC) over
F1 score and accuracy in binary classification evaluation. BMC Genomics, 21(1), 6.
6. CRISP-DM Consortium. (2000). CRISP-DM 1.0: Step-by-step data mining guide. Brussels: SPSS Inc.
7. Dua, D., & Graff, C. (2019). UCI Machine Learning Repository. Irvine, CA: University of California,
School of Information and Computer Science.
8. Fayyad, U., Piatetsky-Shapiro, G., & Smyth, P. (1996). From data mining to knowledge discovery in
databases. AI Magazine, 17(3), 37–54.
9. Han, J., Pei, J., & Kamber, M. (2022). Data Mining: Concepts and Techniques (4th ed.). Burlington, MA:
Morgan Kaufmann.
www.rsisinternational.org
Page 823
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
10. International Energy Agency (IEA). (2024). Africa Energy Outlook 2024. Paris: International Energy
Agency.
11. Khan, R., Patel, S., & Sharma, P. (2022). Predicting customer dissatisfaction in electricity distribution
companies using machine learning techniques. International Journal of Electrical Power & Energy
Systems, 137, 107814.
12. Kotu, V., & Deshpande, B. (2019). Predictive Analytics and Data Mining: Concepts and Practice with
RapidMiner. Morgan Kaufmann.
13. Loureiro, S. M. C., Romero, J., & Bilro, R. G. (2021). Customer satisfaction and predictive analytics: A
systematic review. Journal of Business Research, 122, 805–816.
14. Mitchell, T. M. (1997). Machine Learning. New York: McGraw-Hill.
15. Nigerian Electricity Regulatory Commission (NERC). (2023). Customer Service Standards of
Performance for Distribution Companies. Abuja: NERC.
16. Nigerian Electricity Regulatory Commission (NERC). (2024). Annual Report and Industry Statistics.
Abuja: NERC.
17. Nigerian Electricity Regulatory Commission (NERC). (2025). Quarterly Report on Customer
Complaints and Resolution Statistics. Abuja: NERC.
18. Nwankwo, D., Eze, C., & Okafor, U. (2022). Application of machine learning algorithms in customer
satisfaction prediction within utility sectors. Journal of Big Data, 9(1), 88.
19. Okoye, C. N., & Nwachukwu, K. C. (2022). Assessment of customer satisfaction with electricity service
delivery in Southeastern Nigeria. Energy for Sustainable Development, 67, 72–80.
20. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., ... Duchesnay, E. (2011).
Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12, 2825–2830.
21. Provost, F., & Fawcett, T. (2013). Data Science for Business. Sebastopol, CA: O'Reilly Media.
22. Rani, P., Gupta, S., & Sharma, M. (2021). Comparative analysis of machine learning algorithms for
customer satisfaction prediction using structured and unstructured datasets. Expert Systems with
Applications, 168, 114324.
23. Singh, A., & Kumar, R. (2021). Using data mining techniques to predict customer satisfaction in utility
companies. Expert Systems with Applications, 176, 114846.
24. Witten, I. H., Frank, E., Hall, M. A., & Pal, C. J. (2023). Data Mining: Practical Machine Learning Tools
and Techniques (5th ed.). Morgan Kaufmann.
25. World Bank. (2024). Tracking SDG7: The Energy Progress Report 2024. Washington DC: World Bank.
26. Zhang, X., Li, Y., & Wang, H. (2022). Predictive analytics for customer behavior forecasting in utility
services using ensemble learning models. Applied Energy, 314, 118987.
