In today’s world of design advancements, there is a growing need for energy-saving lighting setups to cut costs and promote eco-friendliness. As per the data from the International Energy Agency (IEA), lighting contributes to a considerable portion of the energy usage worldwide [1]. Conventional indoor lighting setups typically follow set timetables or manual adjustments without considering the changing daylight and the occupancy trends in the space. Excessive energy usage due to inefficiencies can disrupt the comfort of occupants by relying on artificial lighting even when natural daylight is sufficient. This not only leads to energy waste but also causes inconvenience for users [2].

In addition, energy-saving and higher human comfort have been achieved in buildings with efficient luminance using intelligent lighting systems that respond immediately to variation in ambient light [3]. However there are still some lighting setups in intricate spaces that are not as efficient. Too often, they operate independently of daylight levels and environmental changes which are also difficult to scale up and adopt.

Different methods have been explored to improve lighting conditions in research projects. For instance a research conducted by Vashishtha and colleagues introduced a fuzzy controller that uses reinforcement learning to modify lighting settings according to user responses.However this reliance on user interaction could potentially cause

user exhaustion in the long run. Moreover, the study is entirely user choice based and does not consider external parameters such as the natural light variation or people occupancy which are critical to establish the proper lighting. [4]. Also Basurto and co showed that the real time optimization can be realized with small embedded devices such, as Raspberry Pi. But they had difficulties scaling the system[5].

Putrada et al presented a cloud-based control system using a neural network to control indoor conditions with a focus on reducing energy usages and lighting level. However, in such environments, the application of numerous and varied sensor networks can be an expensive and intricate solution [3]. While somewhat similar, a proposed approach by Xie and Sawyer tried to use minimized glare and passive energy performance examples led by simulated daylight data to enhance energy efficiency and glare controls in buildings, and may encounter the difficulty to adopt to real time changes[6]. Moreover, Tm, Kurian and Shetty cited unifying machine learning with lighting and automated blinds but then ran into expansion problems caused by the complexity of managing all the built-in systems in an efficient mode [7].

Addressing these obstacles involves aiming to improve energy efficiency and increase satisfaction through the deployment of an advanced lighting control system that incorporates efficient tools,like embedded platforms (ESP32), the Internet of Things (IoT) and machine learning (ML). By leveraging these technologies and integrating ML into the energy management systems increases the accuracy of adjusting the desired light intensity by taking into consideration and train on a variety of factors, including real-time and historical data [8]. Decision Trees and Neural Networks are the best choices for implementing such systems due to many factors, such as dealing with non-linearity and being light weight models.[5 ,8]. Additionally, using IOT can provide seamless data communications between different devices and remote control of the system, which leads to improved adaptability [9]. Moreover, Microcontrollers like ESP32 have one or dual Tensilica Xtensa 32-bit LX6 microprocessor(s), which enable them to achieve complex tasks such as implementing ML methods on the edge, with some constraints due to their light specs [10]. These methods can be utilized to run ML models without the need for using a cloud connectivity, which in turn, leads to a decentralized system with quick responsiveness and reliability [11].

In short this study introduces a lighting system that automatically regulates brightness levels based on sensor inputs and predictive algorithms to accommodate variations in surroundings and human activity. By merging cloud connected monitoring resources with energy conserving techniques the system gains scalability and user friendliness. This unique method utilizes technologies to address the limitations of current systems resulting in enhanced energy efficiency and improved comfort, for those inside.

2.1          System Design

The system design is based on a decentralized network of nodes. Each node is composed of a processor (ESP32) connected to two sensors and RTC and MOSFET modules. The LDR and PIR sensors measure the ambient light level and human occupancy, while the MOSFET module generates the desired signal for adjusting the LED light intensity. The RTC module creates the system’s timestamp. Additionally, the ESP32 uses its Wi-Fi connectivity to retrieve weather data through the OpenWeather API. All these data are used as inputs for training different machine learning models to predict and adjust the LED light brightness.

Figure 1 shows the overall system architecture, illustrating the three-node setup where each node communicates via a Wi-Fi network with cloud services for data exchange and weather integration.

2.2          Hardware Components

2.2.1       Power Supply and Buck Converters

As shown in Figure 2, the LED driver, which is embedded with the LED light, converts the AC main power supply (220V) to 12V DC to feed the LED light directly and the system through a buck converter. Subsequently, the buck converter steps down the 12V DC to 5V DC to supply the ESP32 along with the sensors to ensure the system’s proper functioning without using a backup battery except for the RTC module, which was explained earlier [12].

Figure 1: Block Diagram of the Overall System Architecture, illustrating three identical nodes, each responsible for controlling one light, and their communication with cloud services.

1. ESP32 Microcontroller

The ESP32 microcontroller, developed by Espressif Systems, is equipped with a dual-core CPU and integrated Wi-Fi capabilities [13]. It serves as the main processing unit, handling sensor data, executing machine learning algorithms, and generating Pulse Width Modulation (PWM) signals to control LED brightness. Its ability to manage data processing tasks makes it suitable for edge computing applications [11].

1. Real-Time Clock (RTC) Module

The DS3231 RTC module provides accurate timekeeping for scheduling lighting adjustments and recording system data [14]. In addition, the module’s battery can maintain the time and correct timestamp, which allows the system to reliably operate even in power outages [14].

1. Sensors

Light Dependent Resistors (LDRs) LDR sensors measure ambient light intensity, providing input for deter- mining the need for artificial lighting adjustments [15]. They are cost-effective and widely used in smart lighting systems to enhance energy efficiency [16].

Passive Infrared (PIR) Motion Sensors PIR sensors detect occupancy by sensing infrared radiation from moving objects [17]. This information is used to adjust lighting based on the presence of occupants, ensuring energy is not wasted in unoccupied spaces [16].

1. LED Lighting and Drivers

Dimmable LED bulbs (40W equivalent, approximately 800 lumens) emitting warm white light at 2700K were used [18]. The brightness is adjustable from 0% to 100% using PWM signals from the ESP32, facilitated by MOSFET transistors acting as electronic switches [19]. An LED driver converts the 220V AC supply to 12V DC to safely power the LEDs [20].

The integration of these hardware components enables the system to collect data, process information, and control the lighting environment effectively. Understanding how these components interact is crucial for the implementation of the control mechanism.

1. Control Mechanism

When motion is detected using the PIR sensor, the lights maintain the predetermined brightness level. If no motion is detected for 2 minutes, the lights dim to 5% brightness for safety [16]. The LED brightness is controlled using PWM signals, with the duty cycle D calculated as:

D = Ldesired          100%

Lmax

where L_desired is the predicted light intensity from the machine learning model, and L_max is the maximum brightness.

Figure 2: Complete Connection Circuit Diagram of the Adaptive Lighting Control System, detailing the connections between ESP32 microcontrollers, sensors, MOSFETs, LED drivers, and power supply components.

1. Software Components

   1. Machine Learning Models

      • Decision Tree Regression and Multi-Layer Perceptron (MLPs): These models were built using Python on Google Colab. Decision Trees were chosen for their ability to handle non-linear relationships and ease of interpretation [8]. MLPs,on the other hand, were selected for their capability to capture complex interactions between features [21].The Decision Tree model was transformed into header files and integrated into C++ code within the Arduino IDE for use on the ESP32 microcontroller [22]. Conversely after training the MLP model, a model file was generated and deployed to the ESP32 using the Edge Impulse platform. This approach was necessary because the TensorFlow library required for the MLP model is incompatible with the ESP32 in Arduino IDE.

1. IoT Platform and APIs

   • Blynk IoT Platform: Employing the BlynK platform facilitated overseeing and managing the lighting system remotely through mobile app and a web-based dashboard [23]. This platform offers a user interface for individuals to engage with the system effectively [9]. Features include a Gauge Widget displaying the current LED brightness percentage (0% to 100%), a Slider Widget for manual brightness adjustment, and a Switch Button to toggle the lights on or off. This integration allows users to rely on automated adjustments or manually control the lighting based on their preferences [9]. Figure 3 shows the Blynk web dashboard and mobile application interfaces in edit modes, respectively.

   • OpenWeatherMap API: This API supplies real-time weather data, enhancing the precision of lighting adjustments by accounting for current weather conditions that affect ambient light levels [24].

1. Experimental Setup

The system was deployed in various locations within a controlled office environment measuring 5 meters by 5 meters. The exact placement of the system depended on the position of the lights within the environment. Data.

1. Blynk Web Dashboard Interface in Edit Mode Showing Gauge, Slider, and Switch Widgets

(b)       Blynk Mobile Application In- terface Displaying Light Intensity Gauge and Control Widgets

Figure 3: Blynk Interfaces for Remote Monitoring and Control in Edit Mode

collection occurred over a period of 5 days, each day for 10 hours under varying lighting conditions and occupancy patterns. Sensors were calibrated before deployment to ensure accurate readings [16].

2.6          Data Collection Procedures

Sensor data was collected at a sampling rate of 1 record per minute. The ESP32 logged ambient light levels from the LDR sensor, occupancy status from the PIR sensor, time from the RTC module, and weather data from the OpenWeather API. Data was transmitted via Wi-Fi to a computer on the same network for storage [10]. After sufficient data was gathered, it was used for analysis and training of machine learning models [8].

Figure 4 illustrates the data collection procedure, showing the flow of data from sensors to storage for analysis and model training.

Figure 4: Data Collection Procedure, illustrating the flow of data from sensors to storage for analysis and model training.

1. Label Optimization

An optimization-based labeling approach was used to generate optimal lighting labels. An objective function balancing occupant comfort and energy efficiency was formulated:

Maximize f (x) = α ×C(x) −β × E(x)

where x is the lighting intensity (0% to 100%), C(x) is the comfort function, E(x) is the energy consumption function, and α and β are weighting coefficients.

The comfort function (C(x)) is defined as:

C(x) = O× (k₁× x−k₂ × L)

and the energy consumption function E(x) is:

E(x) = k₃ × x

In above equations,O represents the occupancy level within the space, while k1 and k2 are constants that influence how lighting intensity x and the baseline light level L affect occupant comfort. Additionally, k3 is the coefficient for the rate of energy consumption.Optimization was performed using Python’s SciPy library, with the optimal intensities serving as target values for training the machine learning models. Figure 5 presents a flowchart of the label optimization process.

Figure 5: Flowchart of the Label Optimization Process, showing the logical steps of collecting data, defining the objective function, optimizing labels, and refining parameters until the optimized labels 
are accepted.

1. Data Processing and Feature Engineering

Our machine learning pipeline, illustrated in Figure 6, begins with a dataset that includes optimized lighting intensity labels designed to balance occupant comfort and energy efficiency. In the data preprocessing stage, we addressed missing values by replacing numerical gaps with the mean and categorical gaps with the mode, ensuring a complete and consistent dataset [25]. We then normalized the numerical features using Min-Max scaling to standardize their range between 0 and 1, facilitating more effective model training [11]. Additionally, categorical variables were transformed into binary vectors through one-hot encoding, which are simpler for machine learning models to process.

After preparing the data for analysis, we made some adjustments to improve the dataset’s ability to predict outcomes. In the first step, 3 new features were included in the optimized dataset: the rate of ambient light changes to capture dynamic fluctuations, moving averaged occupancy rates to smooth out short-term fluctuations and emphasize patterns, and combinations of light intensity and time to represent temporal connections [26]. This is called 3-feature engineering. Additionally, 15 sets of features were added, which included interaction terms and contextual elements, like weather conditions, to give the models a broader range of information [27]. This is a 15-feature engineering dataset.

We proceeded training our machine learning models —the Decision Tree and MLP Neural Network— utilizing the preprocessed data in three different scenarios. One without any feature, then with three customized features and a third incorporating fifteen customized features. This process allowed us to evaluate the impact of feature engineering on our model performance. [8].

Figure 6 6 illustrates the flow of the process beginning with the optimized raw data and progressing through preprocessing steps, then operating on enhancing features, and finally training models.

Figure 6: Machine Learning Pipeline for Adaptive Lighting System illustrating the steps from data preprocessing, feature engineering, model training, and evaluation.

1. Performance Metrics

For measuring the performance of the machine learning models, we used these three metrics: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R-squared (R²).

MAE calculates the average absolute deviation from the real values, and is given by the following expression:

RMSE measures the Root mean square error between the predicted and actual values in the following way:

R² is an indication of how well the model explains the variability in the data:

where n is the number of observations, yi is the actual value, yˆ_i is the predicted value, and y is the mean of the actual values. A higher R² value signifies a better fit of the model to the observed data.

1. Energy Efficiency Analysis

The energy saving performance was then evaluated in a five-day test under the office environment. For each of three decentralized systems there was a 40 watt LED bulb connected that could consume overall a total 1.20 kilowatt hours a day if all three were active on full brightness for all testing hours.

The lighting schedule spanned 10 hours a day (8 am to 6 pm). In the non-adaptive system, the lights remained at full brightness throughout the 10-hour period, consuming a total of 6 kWh over five days. In contrast, the adaptive system adjusted light levels based on real-time data from ambient light sensors, occupancy detectors, and weather information [3]. This dynamic adjustment resulted in an average daily consumption between 0.76 and 0.80 kWh, totaling 3.91 kWh over five days and yielding an average energy saving of 34.8%.

Table 1 summarizes the daily energy consumption of both systems:

Table 1: Daily Energy Consumption of Adaptive vs. Non-Adaptive Systems.

As shown in Table 1, the adaptive system consistently reduced energy consumption compared to the non-adaptive system. Figure 7 visually compares the daily energy usage and savings percentage.

Figure 7: Comparison of Daily Energy Consumption and Energy Savings Percentage

The reduction in energy usage is attributed to the system’s ability to regulate light brightness in real-time based on environmental conditions and occupancy patterns. When natural light was sufficient or fewer occupants were present, the system reduced artificial lighting to prevent unnecessary energy consumption, aligning with findings by Ramadhani et al. [16]. For instance, on Day 3, the system achieved the highest energy savings of 36.7%, likely due to optimal natural lighting conditions that allowed for more effective dimming of artificial lights.

The outcomes show how well the adaptive lighting control system improves energy efficiency by adjusting lighting levels based on variable conditions to save energy while keeping occupants comfortable over five days in various operational situations.

In comparison to research findings by Ramadhani et al., [16] who noted a 30 percent decrease in energy usage with a smart room lighting configuration and Cabezas et al., [28] who witnessed savings between 25 and 40 percent, in public smart lighting initiatives. Our systems average energy savings of 34.38% closely match these outcomes showcasing its competitiveness and real world utility.

3.2          Machine Learning Model Performance

To predict optimal lighting intensity, we developed and evaluated two machine learning models: Decision Tree Regression and Multi-Layer Perceptron (MLP) Neural Network. These models were assessed under three

different feature engineering scenarios to determine the impact of additional features on their predictive performance [8].

3.2.1       Baseline Performance (Without Feature Engineering)

In the baseline scenario, where no feature engineering was applied, each model was trained using only the optimized raw input features. The performance metrics for this state are presented in Table 2.

Table 2: Baseline Performance of Machine Learning Models (Without Feature Engineering)

As shown in Table 2, the Decision Tree Regression model exhibited a Mean Absolute Error (MAE) of 36.13, a Root Mean Squared Error (RMSE) of 51.51, and an R2 score of 0.69. The MLP Neural Network showed similar performance, indicating moderate predictive capability. These metrics suggest that while the models capture some underlying patterns in the data, there is significant room for improvement.

3.2.2      Impact of Feature Engineering

Feature engineering has an essential role in improving the performance of machine learning models by transforming raw data into more informative features [29]. We investigated the impact of feature engineering by evaluating the models under two additional scenarios: with 3 engineered features and with 15 engineered features.

Table 3: Performance Metrics of Machine Learning Models Under Different Feature Engineering Scenarios.

Table 3 provides an overview of the performance metrics across different feature engineering scenarios. Introducing three engineered features led to substantial improvements in all models. Specifically, the Decision Tree Regression model’s MAE decreased from 36.13 to 19.25, and its RMSE reduced from 51.51 to 29.10, with the R2 score increasing from 0.69 to 0.89. This enhancement underscores the effectiveness of feature engineering in improving model performance [8], [26].

Expanding the feature set to fourteen engineered features yielded slight additional improvements. Both the Decision Tree Regression and MLP Neural Network models achieved MAE values around 19 and R2 scores of 0.91. These marginal gains suggest that while more features can enhance performance, the benefits may diminish beyond a certain point due to potential overfitting or redundancy [11]. Figure 8 illustrates the performance improvements when transitioning from no feature engineering to 3 and 15 engineered features, highlighting the reduction in MAE and increase in R2 scores.

3.2.3       Visualization of Predicted vs. Actual Values

Figure 9 visualizes the predicted versus actual lighting intensity values for the Decision Tree Regression and MLP Neural Network models under different feature engineering scenarios. Without feature engineering, the models show a wider spread of data points away from the ideal line (where predicted equals actual), indicating less accurate predictions. After applying feature engineering, the data points cluster more closely around the ideal line, demonstrating improved model accuracy.

(a) Mean Absolute Error (MAE)

Figure 8: Comparison of MAE and R2 Scores of Decision Tree Regression and MLP Neural Network Models Under Different Feature Engineering Scenarios.

(c)MLP Neural Network without Feature Engineering.

1. (d) MLP Neural Network with Feature Engineering

Figure 9: Predicted vs. Actual Lighting Intensity Values for Machine Learning Models Under Different Feature Engineering Scenarios.

1. Cross-Validation Analysis

After analyzing the performance data and conducting validation assessments of various models, it was found that the Decision Tree Regression model is the best choice for implementing in the adaptive lighting system. It demonstrated predictive capabilities with an R² of 0.91 and operated efficiently on the ESP32 microcontroller due to its lightweight design[11]. The R² scores of the five folds are represented in Figure 10 for the Decision Tree Regression model. The nearly-identical R² scores across folds corroborate the performance observed earlier on the training and testing splits, meaning the gains made through our feature engineering can be considered robust.

Figure 10: Decision Tree Regression Cross-Validation R² Scores Across Folds

1. Model Selection for Deployment

From the performance data overview and validation of multiple models, the Decision Tree Regression model was identified as the relative best model to choose for implementation in the adaptive lighting system. It had a R²=0.91 predictive performance and worked properly on the ESP32 microcontroller [11]. This balance between accuracy and resource utilization makes it best fit for real time implementations in resource constrained environments.

1. IoT Platform Performance: Blynk Integration

The integration of the Blynk IoT platform enhanced the adaptive lighting control system by enabling remote monitoring and manual control functionalities. As depicted in Figures 11a and 11b, the Gauge widget accurately displayed LED brightness levels, allowing users to monitor real-time lighting conditions. Additionally, the Slider widget enabled users to manually adjust the brightness, offering flexibility beyond the system’s automatic machine learning-based predictions. The Switch button provided a straightforward method for toggling the LED lights on or off, ensuring users could directly control the lighting environment as needed. 

1. Blynk Web Dashboard Interface showing Gauge, Slider, and Switch Widgetsin Active Mode

   1. Occupant Comfort Evaluation

(b)          Blynk Mobile Application In- terface displaying Light Intensity Gauge and Control Widgets in Ac- tive mode.

Figure 11: Blynk Interfaces for Remote Monitoring and Control in Active Mode

3.4          System Responsiveness and Adaptability

The system’s responsiveness was monitored over a 20-hour period, divided into two 10-hour sessions on separate days. The system effectively adjusted lighting levels with response times as fast as 500 milliseconds for instanta- neous changes in ambient light and about one second for occupancy detection. These rapid adjustments ensured that lighting changes were seamless and unnoticeable to users. Figure 12 illustrates the system’s ability to modulate lighting intensity in real-time based on detected lighting levels. The predicted and actual light intensity values closely matched, demonstrating the system’s precision in maintaining desired
illumination levels.

Figure 12: System Responsiveness and Adaptability Over a 20-Hour Period

The adaptive adjustments prevent over-illumination and reduce energy waste. For example, when a cloud passes over the building, reducing natural light, the system promptly increases artificial lighting to maintain consistent illumination levels. Conversely, when natural light increases, the system dims artificial lighting accordingly. This real-time responsiveness aligns with user comfort and energy efficiency goals.

Indirect evaluations hint at occupant contentment, despite not directly gauging comfort levels themselves in this conducted study. Feedback received informally from occupants indicates contentment with the lighting standard maintained by the system that has been effective in minimizing eye strain and enhancing comfort with a consistent deviation of 5 lux. This feedback suggests that the system effectively manages both energy efficiency and the comfort of occupants to establish a work environment for all individuals involved.The consistency in lighting levels and ability to adapt to fluctuations play a key role, in improving visual comfort and productivity levels.

This study has effectively created an expandable indoor lighting management system using ESP32 microcontrollers and IoT technologies with machine learning algorithms to enhance energy efficiency and occupant satisfaction levels which significantly improved by monitoring ambient light and occupancy levels using LDR and PIR sensors and integrating external weather information from the OpenWeatherMap API for precise lighting adjustments. The machine learning models, in particular Decision Tree Regression was able to predict fairly well with R² scores of 0.91 and higher. The optimized models were deployed to ESP32 microcontrollers, which are very limited in resources but can provide IoT real-time smart lighting control solutions. Experimental results showed a 34.8% energy savings over non-adaptive systems, validating the performance of the dynamic data-driven approach. The integration of the Blynk IoT platform enabled monitoring and control by users to improve the systems usability and adaptability. The decentralized node setup ensured scalability for settings ranging from small offices to large commercial spaces. The quick response and ability to adjust to changes also helped maintain ideal lighting levels and comfort, for occupants.

FUTURE WORK

Despite these achievements, there are areas that can be improved upon in the future. Enhancing the systems capabilities by adding environmental sensors, like temperature and humidity sensors could help in creating a more thorough environmental monitoring system. By incorporating user preferences and feedback mechanisms we could potentially customize lighting settings to better suit occupants needs and increase their satisfaction. Exploring advanced machine learning models, such as deep learning algorithms might enhance predictive accuracy although we would have to consider the computational limitations of the ESP32 platform. By incorporating energy saving federated learning methods we can improve system efficiency without compromising data confidentiality [30]. Another crucial aspect is bolstering security using smart packet filtering and machine learning techniques [25], [29]. Conducting long term studies, in various real world environments would be beneficial to evaluate scalability and uncover issues linked to widespread implementations [31].

Acknowledgments: Not Applicable

Conflict of Interest Statement: The authors declare that there are no conflicts of interest regarding the publication of this manuscript.

The authors declare that they have no conflict of interest

No funding sources

The study was approved by the University of Information Technology and Communications, Baghdad, Iraq

1. IEA. "Lighting," International Energy Agency, www.iea.org/energy-system/buildings/lighting, Accessed 27 Apr. 2024.

2. Chen, Z., et al. "IoT-Based Smart and Intelligent Smart City Energy Optimization," Sustainable Energy Technologies and Assessments, 49(2022) p. 101724. https://doi.org/10.1016/j.seta.2021.101724.

3. Putrada, A. G., et al. "Machine Learning Methods in Smart Lighting Toward Achieving User Comfort: A Survey," IEEE Access, 10(2022) pp. 45,137–45,178. https://doi.org/10.1109/ACCESS.2022.3169765.

4. Vashishtha, K., et al. "Reinforcement Learning Adaptive Fuzzy Controller for Lighting Systems: Application to Aircraft Cabin," arXiv Preprint, arXiv:2310.00525, 2023. https://doi.org/10.48550/arXiv.2310.00525.

5. Basurto, C., et al. "Implementation of Machine Learning Techniques for the Quasi Real-Time Blind and Electric Lighting Optimization in a Controlled Experimental Facility," Journal of Physics: Conference Series, 2042(1)(2021) p. 012112. https://doi.org/10.1088/1742-6596/2042/1/012112.

6. Xie, J., and A. Sawyer. "Simulation-Assisted Data-Driven Method for Glare Control with Automated Shading Systems in Office Buildings," Building and Environment, 196(2021) p. 107801. https://doi.org/10.1016/J.BUILDENV.2021.107801.

7. Kumar, T. S., et al. "A Data-Driven Approach for the Control of a Daylight–Artificial Light Integrated Scheme," Lighting Research & Technology, 52(2)(2020) pp. 292–313. https://doi.org/10.1177/1477153519841104.

8. Ghazal, T. M. "Energy Demand Forecasting Using Fused Machine Learning Approaches," Intelligent Automation & Soft Computing, 31(1)(2022) pp. 539–553. https://doi.org/10.32604/iasc.2022.019658.

9. Devi, M., et al. "Design and Implementation of a Smart Home Energy Management System Using IoT and Machine Learning," E3S Web of Conferences, 387(2023) p. 04005. https://doi.org/10.1051/e3sconf/202338704005.

10. Dokic, K. "Microcontrollers on the Edge – Is ESP32 with Camera Ready for Machine Learning?" Image and Signal Processing, edited by A. El Moataz et al., Springer, 2020, pp. 213–220. ISBN 978-3-030-51935-3.

11. Jouini, O., et al. "A Survey of Machine Learning in Edge Computing: Techniques, Frameworks, Applications, Issues, and Research Directions," Technologies, 12(6)(2024) https://doi.org/10.3390/technologies12060081.

12. Mokhtar, H., et al. "IoT-Based Monitoring of Refrigerated Vaccine Storage: A Literature Review and a Proposed Solution," ICCIT (2023) pp. 20–23. https://doi.org/10.1109/ICCIT58132.2023.10273872.

13. Espressif Systems. "ESP32 Series Datasheet," www.espressif.com/en/products/socs/esp32, Accessed 27 Apr. 2024.

14. Raj, S. V., et al. "Real-Time-Clock Using Arduino," UGC Care Gr. I J, 13(5)(2023) pp. 356–363.

15. Susilawati, A., et al. "Light Intensity Distribution in the Room Using Light Dependent Resistor: Through the Engineering Design Process," Indonesian Journal of Science and Technology, 9(3)(2024) pp. 679–708. https://doi.org/10.17509/ijost.v9i3.74410.

16. Ramadhani, R. R., et al. "Smart Room Lighting System for Energy Efficiency in Indoor Environment," International Journal of Artificial Intelligence & Robotics (IJAIR), 4(2)(2022) pp. 48–58. https://doi.org/10.25139/ijair.v4i2.5266.

17. Ergi, B. N. Gerald, and S. Mandala. "PIR Sensor-Based Intelligent Home Security System Design Analysis Using Machine Learning Methods," ICoDSA 2023 (2023) pp. 175–180. https://doi.org/10.1109/ICoDSA58501.2023.10277399.

18. Philips Lighting. "LED Lamps and Tubes," Signify, www.signify.com/global/brands/philips, Accessed 27 Apr. 2024.

19. Malarczyk, M., et al. "Electric Drive with an Adaptive Controller and Wireless Communication System," Future Internet, 15(2)(2023) https://doi.org/10.3390/fi15020049.

20. Atmadja, W., et al. "Indoor Hydroponic System Using IoT-Based LED," IOP Conference Series: Earth and Environmental Science, 998(1)(2022) p. 012048. https://doi.org/10.1088/1755-1315/998/1/012048.

21. Hussein, M., et al. "Smart Cognitive IoT Devices Using Multi-Layer Perception Neural Network on Limited Microcontroller," Sensors, 22(14)(2022) https://doi.org/10.3390/s22145106.

22. Arduino Software. "Arduino IDE," www.arduino.cc/en/software, Accessed 27 Apr. 2024.

23. Blynk Inc. "Blynk IoT Platform," www.blynk.io, Accessed 27 Apr. 2024.

24. OpenWeatherMap. "Weather API," www.openweathermap.org/api, Accessed 27 Apr. 2024.

25. Saba, T., et al. "A Machine-Learning-Based Approach for Autonomous IoT Security," IT Professional, 23(3)(2021) pp. 69–75. https://doi.org/10.1109/MITP.2020.3031358.

26. Du, P., et al. "Faster Deployment for Indoor Visible Light Positioning Using XGBoost Algorithms in Industrial Internet-of-Things," IECON 2021 – 47th Annual Conference of the IEEE Industrial Electronics Society, (2021) pp. 1–7. https://doi.org/10.1109/IECON48115.2021.9589151.

27. Glória, A. F. X., and P. J. A. Sebastião. "Autonomous Configuration of Communication Systems for IoT Smart Nodes Supported by Machine Learning," IEEE Access, 9(2021) pp. 75,021–75,034. https://doi.org/10.1109/ACCESS.2021.3081794.

28. Cabezas, F., et al. "Optimization Solutions in Energy Management Systems for Public Smart Lighting Applications," IEEE EEEIC / I&CPS Europe 2023, (2023) pp. 1–6. https://doi.org/10.1109/EEEIC/ICPSEurope57605.2023.10194635.

29. Bertoli, G. d. C., et al. "Design and Implementation of Intelligent Packet Filtering in IoT Microcontroller-Based Devices," Journal of Internet Services and Applications, 15(1)(2024) pp. 289–301. https://doi.org/10.5753/jisa.2024.3835.

30. Zhang, T., and S. Mao. "Energy-Efficient Federated Learning with Intelligent Reflecting Surface," IEEE Transactions on Green Communications and Networking, 6(2)(2022) pp. 845–858. https://doi.org/10.1109/TGCN.2021.3126795.

31. Zekic-Sušac, M., et al. "Machine Learning-Based System for Managing Energy Efficiency of Public Sector as an Approach Towards Smart Cities," International Journal of Information Management, 58(2021) p. 102074. https://doi.org/10.1016/j.ijinfomgt.2020.102074.