Improved Connected-Mode Discontinuous Reception (C-DRX) Power Saving and Delay Reduction Using Ensemble-Based Traffic Prediction

Abstract

This paper proposes a traffic prediction-based connected-mode discontinuous reception (C-DRX) approach to enhance energy efficiency and reduce data transmission delay in mobile communication systems. Traditional C-DRX determines user equipment (UE) activation based on a fixed timer cycle, which may not align with actual traffic occurrences, leading to unnecessary activation and increased energy consumption or delays in data reception. To address this issue, this paper presents an ensemble model combining random forest (RF) and a temporal convolutional network (TCN) to predict traffic occurrences and adjust C-DRX activation timing. RF extracts traffic features, while TCN captures temporal dependencies in traffic data. The predictions from both models are combined to determine C-DRX activation timing. Additionally, the extended activation approach is introduced to refine activation timing by extending the activation window around predicted traffic occurrences. The proposed method is evaluated using real-world Netflix traffic data, achieving a 20.9% decrease in unnecessary active time and a 70.7% reduction in mean delay compared to the conventional periodic C-DRX approach. Overall, the proposed method significantly enhances energy efficiency and quality of service (QoS) in LTE and 5G networks, making it a viable solution for future mobile communication systems.

1. Introduction

1.1. Background

In modern mobile communication environments, the demand for ultra-fast data transmission and low-latency communication has grown exponentially [1,2]. This surge has been driven by the widespread adoption of data-intensive applications such as video streaming, online gaming, and internet of things (IoT) services. To meet these demands, higher frequency bands and advanced network protocols have been introduced [3,4], leading to a significant increase in energy consumption in user equipment (UE) [5,6].

To mitigate this issue, discontinuous reception (DRX) technology was introduced as a power-saving mechanism that deactivates the receiver during idle periods, thereby reducing power consumption [7,8,9]. Particularly, connected-mode DRX (C-DRX) extends DRX functionality to the RRC_CONNECTED state, allowing UEs to periodically enter a sleep state to conserve energy [10].

According to the third generation partnership project (3GPP) standards, DRX parameters are primarily determined by the access point (AP) and configured for each UE. However, considering the dynamic nature of power-saving and low-latency requirements, 3GPP has standardized the signaling process of UE Assistance Information in the 5G new radio (NR) system [11,12]. This allows UEs to communicate their preferred DRX parameter settings to balance power saving and DRX-induced latency. For instance, UEs prioritizing power saving (e.g., massive IoT applications or battery-constrained devices) may prefer longer DRX cycles, while UEs requiring low latency (e.g., extended reality (XR) applications or mission-critical IoT services) may opt for shorter DRX cycles [13].

However, conventional C-DRX mechanisms have inherent limitations. DRX parameters are primarily controlled by predefined timer values, causing UEs to alternate between active and sleep states at fixed intervals. While 3GPP has introduced signaling-based adjustments through UE Assistance Information, this approach remains reactive, as adjustments are made only after traffic variations are detected and reported. Due to these limitations, the system lacks the ability to proactively adjust DRX cycles, making it difficult to respond immediately to traffic fluctuations. As a result, when a sudden traffic burst occurs, the UE may remain in sleep mode and fail to receive incoming data in a timely manner. A more detailed discussion of existing C-DRX mechanisms is provided in Section 2.3.

1.2. Related Work

Recent studies have explored various approaches to improve DRX activation and enhance power efficiency.

In [14], a DRX scheduling method based on deep reinforcement learning (DRL) was investigated. This paper proposed dynamically adjusting the wake-up window length and modeled the DRX mechanism as a Markov Decision Process (MDP). A reward-based learning approach was utilized to optimize DRX state transitions, leading to improved power efficiency compared to conventional methods. However, the paper also highlighted a major drawback: the high computational cost associated with training and deploying the DRL model. DRL-based methods require extensive simulations and iterative learning, making them less practical for real-time mobile network environments.

To address this issue, this paper proposes an AI-driven approach utilizing a combination of RF and TCN to analyze traffic patterns and determine DRX activation timing. Unlike DRL-based methods, the proposed approach does not rely on reward-based iterative learning, significantly reducing computational overhead while maintaining adaptive scheduling capabilities.

In [15], an alternative approach was explored by incorporating Downlink Control Information for Power saving (DCP) as a Wake-Up Signal (WUS) to increase the DRX sleep ratio in multi-service environments. This method employed a discrete Markov chain model to adjust DRX scheduling based on service-specific latency requirements. The simulation results demonstrated that this mechanism improved the DRX sleep ratio by up to 10% compared to conventional DRX settings. However, since the DCP-based mechanism depends on explicit wake-up signaling, it is inherently constrained to predefined service scenarios and lacks adaptability to unpredictable traffic patterns.

In contrast, the proposed DRX activation method leverages traffic prediction to determine wake-up timing without requiring additional wake-up signaling (DCP). By eliminating the dependency on explicit wake-up signals, this method ensures adaptability across diverse traffic conditions while minimizing control signaling overhead.

1.3. Contribution

This paper proposes a traffic prediction-based C-DRX activation adjustment method to address the limitations of fixed activation cycles, which often lead to unnecessary energy consumption and increased data reception latency. To overcome this, we introduce an ensemble model combining random forest (RF) and temporal convolutional network (TCN) to predict traffic occurrences and dynamically adjust C-DRX activation timing.

The main contributions of this paper are as follows:

• Traffic-aware C-DRX activation: Adjusting wake-up timing based on predicted traffic patterns to reduce unnecessary UE activations and enhance scheduling efficiency.
• Performance improvement: Experimental results show that the proposed method reduces cumulative active time by 20.9% and decreases mean data reception delay by 70.7%, significantly improving energy efficiency and network responsiveness.
• Computational efficiency: The RF-TCN model minimizes computational complexity while maintaining high predictive accuracy, making it suitable for real-time applications.
• Real-world applicability: The proposed method is evaluated on Netflix traffic data, demonstrating its effectiveness in practical mobile networks.

The remainder of this paper is organized as follows. Section 2 introduces the system model, including the network architecture and C-DRX operation principles. Section 3 describes the characteristics of the traffic dataset, introduces the extended activation approach to mitigate temporal misalignment in traffic predictions, defines key performance metrics for evaluating prediction accuracy, and outlines the methodology for adjusting C-DRX timer parameters. Section 4 presents the proposed traffic prediction model, detailing the individual application of random forest (RF) and temporal convolutional network (TCN), as well as an ensemble approach that combines both models for improved prediction performance. Section 5 analyzes simulation results, including the evaluation of traffic prediction performance, energy consumption, and delay reduction. Finally, Section 6 summarizes the findings of this paper and discusses potential directions for future research.

2. System Model

2.1. Network Architecture

Modern distributed network architectures consist of the datacenter network, core network, access points (APs), and user equipment (UE), forming a hierarchical structure that ensures efficient data transmission and connectivity [16,17,18]. Figure 1 illustrates this architecture.

The datacenter network comprises application servers and cloud computing resources, responsible for large-scale data processing before forwarding data to the core network. The core network connects the datacenter to access points, managing data routing while ensuring ultra-low latency, network stability, and scalability.

APs act as intermediaries between the core network and UEs, facilitating bidirectional data transmission. They relay incoming data from the core network to UEs and transmit user-generated data back to the core network. By enabling localized data processing, APs help reduce transmission latency and improve UE energy efficiency.

UEs include devices such as smartphones, tablets, and IoT devices, which significantly impact network energy efficiency and user experience. To conserve battery life, UEs employ C-DRX, activating only during scheduled reception periods. However, misalignment between activation timing and actual traffic arrivals can lead to increased energy consumption and data reception delays.

To address this issue, this paper proposes an improved prediction-based C-DRX mechanism deployed at the AP. Specifically, an ensemble model combining random forest (RF) and temporal convolutional network (TCN) analyzes historical traffic patterns to predict traffic arrival times. Offloading traffic prediction tasks to the AP reduces computational overhead on UEs, extends battery life, and enhances overall network energy efficiency.

2.2. Traffic Model

According to the Ericsson Mobility Report, mobile network traffic has nearly doubled over the past two years, primarily driven by the increasing consumption of video content. Currently, video traffic accounts for 71% of total mobile data traffic and is projected to rise to 80% by 2028. Video-on-demand (VoD) streaming services such as Netflix, YouTube, and TikTok continue to dominate high-data-consumption user groups, establishing video traffic as a core component of mobile data usage [19]. Reflecting this trend, this paper selects Netflix traffic as a representative case for in-depth analysis.

This paper utilizes a real-world 5G traffic dataset presented in [20] to investigate a traffic prediction-based C-DRX activation mechanism. The dataset consists of publicly available measurement data collected from a major South Korean mobile network operator, covering 328 h of traffic across various applications, including live streaming, stored streaming, video conferencing, metaverse applications, and online gaming. Traffic measurements were conducted using a Samsung Galaxy A90 5G device equipped with a Qualcomm Snapdragon X50 5G modem. Packet capture was performed using PCAPdroid, an open-source packet sniffing tool. To ensure accurate application-specific traffic characteristics, traffic was collected separately for each application, eliminating background interference.

For this paper, Netflix traffic was extracted as the primary analysis target. The extracted dataset includes 24 h and 43 min of Netflix traffic transmitted using the TCP protocol, collected under real user streaming conditions. The total recorded file size is approximately 0.74 GB, providing an authentic representation of actual 5G network conditions rather than simulated traffic patterns.

Conventional C-DRX mechanisms operate at the subframe level (in the order of milliseconds), requiring high temporal resolution. However, large-scale network traffic datasets rarely provide bitrate information at millisecond-level granularity. The dataset used in this paper consists of bitrate data sampled at one-second intervals, making it the most feasible resolution for this paper. Given these practical constraints, this paper utilizes real-world bitrate data recorded at one-second intervals to predict traffic arrival times.

This paper applies a binary conversion method to simplify the learning process. In this approach, traffic presence at a given time is denoted as ‘1’ and its absence as ‘0’, transforming the problem into a binary classification task. Binary conversion offers advantages such as reduced model complexity and enhanced generalization performance. Instead of predicting continuous bitrate values, the model focuses on determining traffic presence, which aligns with the primary objective of C-DRX: minimizing unnecessary power consumption by activating the UE only when data transmission occurs.

Nevertheless, binary conversion may result in the loss of burst traffic characteristics. Netflix traffic, in particular, exhibits sudden bursts of data transmission, which a simple binary representation may fail to capture. For instance, momentary traffic spikes may be represented as continuous sequences of ‘1’s, potentially overlooking subtle variations in traffic patterns.

Despite this limitation, binary conversion remains a suitable choice as this paper primarily aims to refine C-DRX activation timing. Experimental results demonstrate that this approach effectively reduces unnecessary activations and minimizes data reception delays, confirming its practical applicability in energy-efficient DRX scheduling.

Future research will analyze the impact of binary conversion on model performance, specifically assessing how the loss of fine-grained traffic variations affects C-DRX activation accuracy. Additionally, if millisecond-level sampling data become available, further investigations will be conducted to extend the scope of this paper and refine C-DRX activation strategies.

2.3. C-DRX Operation

Discontinuous reception (DRX) is a power-saving mechanism introduced in LTE networks to enhance the energy efficiency of user equipment (UE). While conventional DRX operates only in the RRC_IDLE state, connected-mode DRX (C-DRX) extends this functionality to the RRC_CONNECTED state, allowing UEs to periodically deactivate their receivers while maintaining an active connection. The primary goal of C-DRX is to minimize unnecessary energy consumption by reducing the time the UE monitors the downlink channel. Figure 2 illustrates the timer-based operation of C-DRX, where active and sleep states are controlled by multiple timers [11,12].

C-DRX operation is managed through the following timers:

• onDurationTimer: Defines the initial active period at the start of each C-DRX cycle. During this time, the UE monitors the physical downlink control channel (PDCCH) to detect incoming data transmissions.
• drx-InactivityTimer: Extends the active period following data reception, preventing frequent transitions between sleep and active states.
• shortDRX-Cycle: Specifies a short sleep interval for applications requiring frequent wake-ups, such as video streaming and video conferencing.
• drxShortCycleTimer: Determines the number of short C-DRX cycles before transitioning to a long C-DRX cycle.
• longDRX-Cycle: Defines an extended sleep interval to maximize power savings when traffic is sparse.

The C-DRX state transition process can be mathematically modeled as follows:

$$S\left(t\right)=\left\{\begin{array}{cc}1\hfill & \mathrm{if}{T}_{\mathrm{on}}\left(t\right)>0\mathrm{or}{T}_{\mathrm{inactivity}}\left(t\right)>0\hfill \\ 0\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

(1)

where the following are denoted:

• $S\left(t\right)$: UE state at time t where 1 represents the active state, and 0 represents the sleep state.
• $T_{\mathrm{on}}\left(t\right)$: Active status of the onDurationTimer at time t.
• $T_{\mathrm{inactivity}}\left(t\right)$: Remaining duration of the inactivity timer at time t.

Although conventional C-DRX mechanisms effectively reduce power consumption, they rely on predefined timer values, limiting adaptability to dynamic traffic variations. While network operators can adjust C-DRX timers based on long-term traffic analysis or policy changes, these modifications remain static and cannot respond to real-time traffic fluctuations. As a result, unexpected traffic surges may leave UEs in sleep mode, causing data reception delays and degrading the performance of real-time applications such as VoIP and XR services.

To address these issues, 3GPP has standardized UE Assistance Information in TS 38.300 and TS 38.331, allowing UEs to communicate their preferred DRX configurations to the network [11,12]. This mechanism enables dynamic adjustments based on power-saving preferences and application-specific QoS needs. For example, IoT devices may request longer DRX cycles to extend battery life, whereas latency-sensitive applications such as XR streaming or mission-critical IoT services may opt for shorter cycles. However, the final C-DRX settings are determined by the gNB, which optimizes resource allocation based on network conditions. If the network is congested, shorter C-DRX cycles may not be accommodated, and longer cycles may be overridden for high-traffic UEs to maintain communication quality. Despite these constraints, UE Assistance Information offers greater flexibility than static configurations, balancing power savings and QoS [21].

Nevertheless, UE Assistance Information-based adjustments are inherently reactive, relying on past traffic reports rather than predicting future variations. As a result, sudden traffic surges may still leave UEs in sleep mode when data reception is required, causing delays. To overcome this limitation, this paper proposes an AI-driven C-DRX activation approach. Unlike conventional C-DRX, which reacts based on predefined rules or UE reports, the proposed method leverages machine learning and deep learning techniques to anticipate traffic fluctuations and dynamically refine C-DRX settings.

The proposed AI-enhanced C-DRX adjustment method introduces two key features:

• Proactive wake-up adjustment: The AI model predicts traffic occurrences in advance and fine-tunes C-DRX activation timing to minimize power consumption and reduce data reception delays.
• Traffic pattern adaptation: A hybrid ensemble model combines feature extraction using a random forest (RF) with sequential pattern learning through a temporal convolutional network (TCN), allowing the system to adapt to traffic variations and adjust C-DRX activation timing.

By incorporating traffic prediction into C-DRX management, the proposed method enhances both power efficiency and QoS, addressing the limitations of conventional approaches. The detailed implementation and evaluation of the AI-enhanced C-DRX approach are discussed in Section 5.

3. Traffic Data Characteristics and Baseline Model

3.1. Extended Activation Approach

The traffic data used in this paper are based on real-world network measurements and consist of bitrate information recorded at one-second intervals. The dataset exhibits an imbalanced distribution, with traffic presence accounting for only 8.6% of the total data. This imbalance makes it challenging for prediction models to align perfectly with actual traffic occurrences, leading to temporal misalignment issues.

Prediction models estimate the probability of traffic occurrence at a specific time. However, when actual traffic occurs slightly earlier or later than the predicted time, the UE may remain in sleep mode, resulting in data reception delays. To mitigate this issue, the proposed method extends the C-DRX activation window by $\Delta T$ seconds before and after the predicted traffic occurrence. For example, if the model predicts traffic at time t, the activation period is extended from $t-1$ to $t+1$ seconds. This adjustment reduces the risk of missing traffic due to temporal misalignment and enables more stable traffic handling.

The key idea of this method is to refine C-DRX activation timing based on traffic prediction. The extended activation approach operates as follows: if a predicted traffic occurrence exists at a given time, the corresponding time step remains active. Additionally, the activation window is expanded by $\Delta T$ seconds before and after the predicted occurrence, allowing the UE to receive data at an appropriate time. The original predictions remain unchanged, and the final activation decision follows the extended activation rule.

This process is mathematically expressed as follows:

$$S_{\mathrm{extended}}\left(t\right)=\left\{\begin{array}{cc}1,\hfill & \mathrm{if}{S}_{\mathrm{pred}}(t-\Delta T)=1\mathrm{or}{S}_{\mathrm{pred}}\left(t\right)=1\mathrm{or}{S}_{\mathrm{pred}}(t+\Delta T)=1,\hfill \\ 0,\hfill & \mathrm{otherwise}.\hfill \end{array}\right.$$

(2)

where the following are denoted:

• $S_{\mathrm{pred}}\left(t\right)$: The original traffic occurrence predicted by the model.
• $S_{\mathrm{extended}}\left(t\right)$: The final activation state after applying the extended activation approach.

In this paper, $\Delta T=1$ second is set as the default value, but it can be adjusted based on network conditions and service requirements. Experiments with different $\Delta T$ values confirmed that setting $\Delta T=1$ second provides sufficient performance while minimizing unnecessary activations.

In Figure 3, the green line indicates actual traffic occurrences, the blue line represents the predicted values, and the red line shows the results after applying the extended activation approach.

The original predictions (blue) may sometimes be slightly earlier or later than the actual traffic occurrences (green). However, by applying the extended activation approach (red), the temporal misalignment between prediction and actual traffic is minimized, ensuring that C-DRX activation occurs at more appropriate moments.

Applying the extended activation approach is expected to yield the following benefits:

• Reducing temporal misalignment: Ensures that C-DRX activation better aligns with actual traffic, minimizing data reception delays.
• Enhancing adaptability: Allows dynamic tuning of the activation range based on network conditions and service requirements.

The impact of the extended activation approach on traffic prediction performance is evaluated in Section 4.

3.2. Traffic Prediction Performance Metrics

To evaluate the performance of traffic occurrence prediction, this paper defines three key metrics: OTPa, TOPa, and E-TOPa.

First, the following parameters are defined:

• TP (True positive): The number of correctly predicted traffic occurrences.
• TN (True negative): The number of correctly predicted idle periods.
• FP (False positive): Instances where no traffic was present but incorrectly predicted as occurring.
• FN (False negative): Missed traffic occurrences where actual traffic was present but not predicted.
• ${\mathrm{E}}_{\mathrm{T}}$ (True extended prediction): Additional activations where the extended activation successfully captures actual traffic.

Using these parameters, the performance metrics are defined as follows:

$$\mathrm{OTPa}={\displaystyle \frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}}$$

(3)

$$\mathrm{TOPa}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}}$$

(4)

$$\mathrm{E-TOPa}={\displaystyle \frac{\mathrm{TP}+{\mathrm{E}}_{\mathrm{T}}}{\mathrm{TP}+\mathrm{FN}}}$$

(5)

OTPa measures overall prediction accuracy, considering both traffic occurrences and idle states. In contrast, TOPa quantifies how effectively the model detects traffic occurrences, ensuring that DRX activation is correctly timed.

E-TOPa is specifically designed to assess the effectiveness of the extended activation approach. Unlike standard metrics that evaluate predictions at exact timestamps, E-TOPa accounts for cases where the extended activation successfully aligns DRX activation with actual traffic, reducing errors caused by slight temporal misalignment. A higher E-TOPa indicates better correction for timing discrepancies.

The dataset used in this paper exhibits a significant class imbalance, where idle states (0 s) greatly outnumber traffic occurrences (1 s). In such cases, OTPa alone may not provide a meaningful evaluation, as a model predicting mostly idle states can still achieve a high accuracy score.

To address this issue, this paper uses TOPa to measure the model’s ability to detect actual traffic occurrences, ensuring a more balanced evaluation. Additionally, E-TOPa accounts for minor prediction misalignments, providing a more comprehensive assessment of model performance in imbalanced datasets.

While reducing FN improves data reception, it should not come at the cost of excessive DRX activations, which increase power consumption. Therefore, the challenge is to balance improving traffic prediction accuracy and minimizing unnecessary UE wake-ups.

To achieve this balance, the extended activation approach allows control over the correction strength by adjusting $\Delta T$. A larger $\Delta T$ increases the chance of capturing actual traffic but may lead to unintended activations, whereas a smaller $\Delta T$ conserves power but increases the risk of missing traffic. Selecting an appropriate $\Delta T$ based on the network environment and service requirements is crucial. The extended activation approach provides a flexible framework for refining DRX activation timing.

3.3. Timer-Based C-DRX Parameter Adjustment

The conventional timer-based C-DRX mechanism operates with timers set in milliseconds (ms) to minimize unnecessary power consumption while the UE remains connected to the network. However, the video traffic dataset analyzed in this paper provides bitrate information at one-second intervals, leading to a potential mismatch between the existing C-DRX timer values and the dataset characteristics. If the default C-DRX timer settings are applied without modification, inefficient power management may occur due to misalignment with traffic patterns. Therefore, this paper explores the adjustment of C-DRX timer values based on the characteristics of the traffic data and utilizes Optuna to optimize the parameters through a multi-objective search.

C-DRX consists of multiple timers, each playing a crucial role in balancing UE power savings and data transmission. In this paper, five key parameters were selected for adjustment: onDurationTimer, inactivityTimer, shortSleepTimer, longSleepTimer, and shortSleepRepeatMax. These timers collectively determine the trade-off between power savings and data transmission latency. Therefore, the optimization process focused on three primary performance metrics:

• Average delay: The mean time between traffic occurrence and data transmission.
• Average active time: The proportion of time that the UE remains in an active state.
• False negative rate (FN rate): The proportion of missed traffic occurrences.

To achieve a well-balanced configuration, Optuna was used to perform a multi-objective optimization with 100,000 trials, exploring a wide range of C-DRX timer settings. The optimization process involved the following steps:

• Traffic data processing: The dataset was preprocessed and binarized to distinguish between traffic presence and idle periods.
• Multi-objective optimization with Optuna: Optuna was used to explore a wide range of C-DRX timer settings, minimizing delay, active time, and false negative rate.
• Pareto front selection: The best parameter sets were identified using a Pareto front approach, representing trade-offs where no single metric can be improved without degrading others.
• Balanced solution selection: Among the pareto optimal solutions, the configuration closest to the mean performance values of all three metrics was selected to ensure a well-balanced outcome.

Table 1. Optimized parameters for timer-based C-DRX.

Parameter	Value
onDurationTimer	10 s
inactivityTimer	5 s
shortSleepTimer	22 s
longSleepTimer	220 s
shortSleepRepeatMax	6

summarizes the optimized timer values obtained through this process.

These values are tailored to the specific characteristics of the traffic dataset used in this paper and provide a baseline for evaluating C-DRX performance under different network conditions. Although these values are not universally optimal, they provide a useful reference for evaluating the benefits of dynamic C-DRX adjustment.

4. Proposed Model

4.1. Random Forest

Random forest (RF) is an ensemble learning algorithm that constructs multiple decision trees using randomly sampled subsets of data and features. By aggregating predictions through majority voting, RF mitigates overfitting and enhances generalization performance. It is particularly effective for traffic prediction in C-DRX, as it handles imbalanced data distributions and detects rare traffic occurrences, helping to minimize unnecessary wake-ups. Additionally, its interpretability aids in adaptive C-DRX activation decisions. These characteristics make RF well suited for traffic prediction tasks [22,23].

In this paper, RF was employed to predict traffic occurrences in the Netflix dataset. The dataset was binarized, labeling traffic presence as 1 and absence as 0, formulating the problem as a binary classification task. To incorporate temporal dependencies, a sliding window approach was applied, converting the binarized data into fixed-length input sequences. Experimental evaluations determined that an optimal sequence length of $N=100$ achieved the best performance.

Previous studies have explored RF hyperparameter tuning across a broad range of values. For instance, Ref. [24] examined the impact of n_estimators (100–2000), max_depth (5–100), min_samples_leaf (1–20), and min_samples_split (2–40), showing their significant effect on model accuracy and F1-score. However, applying such broad search spaces in real-world network environments can lead to excessive computational overhead and suboptimal generalization.

To balance model complexity, generalization performance, and computational efficiency, we adopted a narrower but effective search space tailored to the C-DRX traffic prediction task. The defined hyperparameter ranges are as follows:

• n_estimators: 50–300;
• max_depth: 5–30;
• min_samples_split: 2–10;
• min_samples_leaf: 1–10;
• max_features: None, sqrt, log2.

This search space was designed to provide a good trade-off between predictive accuracy and computational feasibility in the context of C-DRX activation timing. To fine-tune these hyperparameters, Optuna was employed for automated optimization, exploring various configurations through an extensive search. The best configuration was selected based on the TOPa metric, maximizing traffic occurrence detection accuracy while minimizing unnecessary UE activations. The optimized hyperparameter values are summarized in

Table 2. Optimized hyperparameters for the random forest model.

Hyperparameter	Value
n_estimators	294
max_depth	19
min_samples_split	4
min_samples_leaf	5
max_features	None

.

As a result, RF maintains robust performance even in highly imbalanced datasets, ensuring reliable OTPa and TOPa scores, making it a strong candidate for traffic prediction tasks in C-DRX management.

4.2. TCN

Temporal convolutional network (TCN) is a deep learning model designed for sequential data processing, employing a stack of causal convolutional layers. Unlike recurrent models, TCN utilizes convolutions to capture temporal dependencies, allowing for efficient parallel processing and stable long-term memory retention [25]. By applying multiple convolutional layers, TCN extracts meaningful features from traffic sequences, making it effective for time-series forecasting.

In this paper, TCN was applied to predict traffic occurrences based on the Netflix dataset. Similar to the random forest model, the dataset was binarized, where traffic presence was represented as 1 and absence as 0. A sliding window approach was used, where a sequence of N previous time steps was utilized to predict traffic occurrence at the next time step. The optimal sequence length was determined as $N=100$ through experimental validation.

The TCN architecture comprises multiple convolutional layers to capture temporal dependencies in traffic sequences. Each layer applies a 1D convolution operation with a fixed kernel size. Batch normalization is applied after each convolution to stabilize training, and ReLU is used as the activation function. The final output is passed through a fully connected layer with a sigmoid activation to classify traffic occurrences.

Optuna was employed for automated tuning to optimize the key hyperparameters of TCN. The hyperparameter search space was designed based on previous studies that explored TCN tuning using various optimization techniques [26]. In particular, prior work utilized Optuna to optimize parameters such as batch size, filter size, kernel size, optimizer type, and learning rate. Building on these findings, we refined the search space to balance predictive accuracy, computational efficiency, and model complexity for C-DRX traffic prediction.

The defined hyperparameter ranges in this paper are as follows:

• filters: 32–128;
• kernel_size: 2–5;
• dense_units: 32–128;
• learning_rate: ${10}^{-4}$–${10}^{-2}$ (log-uniform scale).

This search space was selected to provide a trade-off between accuracy and computational feasibility while ensuring effective traffic prediction for C-DRX activation timing. The optimal hyperparameter values obtained through Optuna optimization are summarized in

Table 3. Optimized TCN hyperparameters.

Parameter	Value
filters	92
kernel_size	3
dense_units	32
learning_rate	0.0034
batch_size	32

.

The model was trained using the Adam optimizer with binary cross-entropy loss. The evaluation was based on TOPa (traffic occurrence prediction accuracy for traffic presence cases).

4.3. Ensemble Model for Traffic Prediction

Ensemble learning is a widely adopted technique that combines multiple models to improve predictive accuracy and stability. In this paper, an ensemble approach is proposed that integrates random forest (RF) and temporal convolutional network (TCN) to enhance traffic prediction performance.

Random forest (RF) and temporal convolutional network (TCN) employ different learning mechanisms, making them complementary for traffic prediction. RF, a decision tree-based model, leverages data randomness to enhance robustness and effectively detect rare traffic occurrences. TCN, a convolutional neural network-based model, captures long-range dependencies in sequential data, making it well suited for modeling temporal patterns in network traffic. By combining these strengths, the ensemble model adapts to diverse traffic patterns beyond the Netflix dataset, ensuring reliable performance across different network environments and application types.

Figure 4 illustrates the structure of the ensemble model. The RF and TCN models are independently trained, and their predictions are combined to generate the final output.

To assess the impact of weighting bias, we experimented with different weighting strategies (e.g., 0.6–0.4, 0.7–0.3). However, experimental results indicated that slight adjustments in weighting did not yield significant improvements in prediction accuracy. Given that RF captures feature-based variations and TCN models sequential dependencies, their equal weighting provided a stable and computationally efficient solution. This approach prevents either model from dominating the prediction while maintaining robust performance. Nevertheless, depending on the dataset characteristics and network conditions, adjusting the weighting scheme could further enhance performance, allowing for greater adaptability in diverse scenarios.

The final classification is determined through a threshold-based binarization process:

$$\mathrm{Final}\mathrm{Prediction}=\left\{\begin{array}{cc}1,\hfill & \mathrm{if}\mathrm{Combined}\mathrm{Prediction}\ge \mathrm{Threshold},\hfill \\ 0,\hfill & \mathrm{otherwise}.\hfill \end{array}\right.$$

(6)

In this paper, the threshold value was set to 0.5, meaning that if the combined prediction exceeds 0.5, the final output is classified as 1 (traffic occurrence), otherwise, it is classified as 0 (no traffic occurrence). By integrating random forest for robust classification and TCN for sequential pattern learning, the proposed ensemble model improves traffic prediction accuracy while maintaining computational efficiency.

5. Simulation Results

5.1. Traffic Prediction Model Performance Comparison

Table 4. Comparison of model accuracy using OTPa, TOPa, and E-TOPa metrics.

Model	OTPa (%)	TOPa (%)	E-TOPa (%)
LSTM	94.9	45.4	68.7
GRU	93.6	26.0	46.0
TCN	95.4	71.2	97.3
RF	96.0	68.0	93.8
GRU + RF	94.6	40.3	71.0
GRU + RF + LGBM	96.0	68.0	91.4
RF + LGBM	96.1	71.6	94.5
RF + XGBoost	95.9	65.8	89.0
LGBM + XGBoost	96.0	67.3	89.7
RF + TCN	95.8	72.6	97.6

compares the performance of different traffic prediction models, including both individual models and ensemble approaches. In this table, the “+” symbol indicates an ensemble model, where multiple models are combined to improve prediction accuracy (see Section 4.3).

As shown in Figure 3, temporal misalignment between actual traffic and model predictions can occur. To mitigate this issue, the extended activation approach was applied (see Section 3.1), resulting in E-TOPa, which reflects the improved performance of TOPa after activation window expansion.

Since network traffic exhibits sequential characteristics, this paper initially evaluated recurrent neural network (RNN)-based models, which are commonly used for processing time-series data. However, as shown in Table 4, the RNN-based models, long short-term memory (LSTM) and gated recurrent unit (GRU), recorded relatively low TOPa values of 45.4% and 26.0%, respectively. Even after applying extended activation (E-TOPa), these models still underperformed. This suggests that while deep learning-based sequential models can be beneficial for traffic prediction, traditional RNN architectures struggled to effectively capture the traffic patterns analyzed in this paper.

In contrast, the temporal convolutional network (TCN), another deep learning-based approach, achieved a significantly higher TOPa of 71.2%, demonstrating its ability to model long-range dependencies in time-series data. Similarly, random forest (RF) recorded a TOPa of 68.0%, highlighting its strength in handling randomness and nonlinearity in traffic data. These results indicate that while traditional RNN-based models were less effective, certain deep learning architectures like TCN and machine learning approaches such as RF can provide strong predictive performance in traffic forecasting.

Although both TCN and RF exhibited strong predictive performance, a single model may struggle to maintain consistent accuracy across diverse traffic conditions and applications. To improve generalizability and stability, this paper adopts an ensemble approach that combines RF and TCN, leveraging their complementary strengths. The RF + TCN ensemble model provides a scalable and adaptable solution for traffic prediction. RF captures key traffic features and handles variations in traffic characteristics, while TCN effectively models sequential dependencies in time-series data. By integrating these capabilities, the ensemble approach extends its applicability beyond the Netflix dataset to diverse traffic scenarios, including web browsing, VoIP, and gaming.

Experimental results confirmed that extended activation improved the RF + TCN ensemble model’s E-TOPa to 97.6%, enhancing stability and reliability in predictions. Among various model combinations, the RF + TCN ensemble exhibited the most balanced performance, achieving the highest E-TOPa. This performance enhancement stems from RF’s ability to detect rare traffic occurrences and TCN’s strength in capturing long-term dependencies, resulting in a well-balanced traffic prediction model.

5.2. Prediction Performance of the Ensemble Model

Figure 5 presents the results of a simulation comparing the performance of the ensemble model and the extended activation approach.

In this paper, we evaluated the models by comparing the predicted traffic occurrence times with actual traffic data. The green line represents actual traffic occurrences, while the blue dashed line and the red dashed line indicate the predictions made by the ensemble model and the extended activation approach, respectively.

As shown in Figure 5, both the ensemble model and the extended activation approach closely align with actual traffic occurrences. The extended activation approach (red dashed line) further improves prediction accuracy by correcting minor temporal misalignments between predicted and actual traffic events.

The ensemble model achieved OTPa of 95.8%, TOPa of 72.6%, and E-TOPa of 97.6%, demonstrating its strong predictive capability. These results confirm that integrating RF and TCN enhances traffic occurrence prediction, while the extended activation approach further refines accuracy by mitigating timing discrepancies.

Moreover, the ability to accurately predict traffic occurrences plays a crucial role in improving C-DRX operational efficiency. By minimizing unnecessary UE activations and reducing energy consumption, the proposed approach enhances network power efficiency and overall system stability.

5.3. Performance Evaluation of the Proposed Model

This paper evaluates the impact of different C-DRX approaches—periodic, prediction-based, and extended prediction-based—on energy efficiency, delay reduction, and traffic detection performance (false negative, FN).

In this paper, cumulative active time is employed as a key metric for evaluating power savings, as a lower cumulative active time indicates fewer unnecessary UE wake-ups, leading to improved energy efficiency. Additionally, FN rates are analyzed to assess the traffic detection performance of each approach.

Figure 6 presents a comparison of cumulative active time across periodic, prediction-based, and extended prediction-based C-DRX approaches.

The periodic C-DRX mechanism continuously increases active time regardless of traffic patterns, as it activates the UE at fixed intervals. In contrast, the prediction-based and extended prediction-based approaches dynamically adjust activation timing by incorporating actual traffic patterns, effectively reducing unnecessary activations.

Experimental results indicate that the prediction-based approach reduces cumulative active time by 61.6% compared to periodic C-DRX, while the extended prediction-based approach achieves a 20.9% reduction. These findings suggest that adjusting C-DRX activation based on traffic prediction can significantly minimize unnecessary UE activation and improve energy efficiency.

Figure 7 illustrates the comparison of mean delay among the periodic, prediction-based, and extended prediction-based C-DRX approaches.

Experimental results show that the prediction-based approach reduces mean delay by 17.9%, while the extended prediction-based approach achieves a 70.7% reduction. Particularly, the extended prediction-based approach recorded the lowest mean delay, ensuring immediate response to incoming traffic and enhancing user QoS.

Figure 8 illustrates the false negative (FN) rate comparison across different C-DRX approaches.

Periodic C-DRX exhibits the highest FN rate at 63.1%, as its fixed activation cycles frequently fail to detect incoming traffic. The prediction-based approach significantly reduces the FN rate by 56.6% (from 63.1% to 27.4%), while the extended prediction-based approach further lowers it by 92.6% (from 63.1% to 4.7%), ensuring more reliable traffic detection.

Experimental results confirm that although the prediction-based approach performs well, the extended prediction-based approach significantly outperforms it in reducing delay and FN rates. However, this improvement comes at the cost of increased active time, leading to higher energy consumption. This trade-off suggests that the prediction-based approach is more suitable when prioritizing energy efficiency, while the extended prediction-based approach is preferable in scenarios where low latency and accurate traffic detection are critical. Thus, selecting the appropriate method depends on the trade-off between energy efficiency and QoS enhancement in a given network environment.

6. Conclusions

This paper proposed an AI-driven traffic prediction-based C-DRX mechanism to improve the efficiency of conventional periodic C-DRX. An ensemble model combining random forest (RF) and temporal convolutional network (TCN) was employed to adjust C-DRX activation. Additionally, an extended activation approach was introduced to refine activation timing and mitigate prediction misalignment. This approach was applied to real-world Netflix traffic data for experimental validation.

The prediction-based approach demonstrated significant energy savings, while the extended activation approach improved traffic detection performance by reducing the false negative (FN) rate and enhancing QoS. These results suggest that the prediction-based approach is suitable for energy-efficient environments, whereas the extended activation approach is more effective in scenarios requiring low latency and high reliability.

Although this paper focused on Netflix traffic, the proposed RF + TCN model, which integrates the strengths of machine learning and deep learning, can be applied to various applications such as web browsing, VoIP, and gaming. Furthermore, it exhibits scalability in handling diverse traffic patterns. Future research will incorporate multiple datasets to evaluate performance under various network conditions and further refine adaptive C-DRX strategies.

In conclusion, this paper highlights the potential of AI-driven traffic prediction in enhancing C-DRX operation for future wireless networks.