Breaking the energy curve: Network energy consumption modeling and energy saving technologies

Network energy consumption is considered a key parameter in designing the 5G New Radio (NR) standard since its inception. This has been motivated by the need to reduce both the carbon footprint of mobile communications and the network operational expenditure (OPEX). Furthermore, the recent rise in energy prices has led to a significant increase in communications service providers' (CSPs’) energy bills. Some CSPs have lowered their profit expectations, while others have increased their use of energy saving functions to save energy costs. It is now more important than ever to break the energy curve.

Ericsson’s report Breaking the energy curve outlines approaches to reducing energy consumption among which modernizing existing networks is particularly crucial. The blog post Crush the energy curve! How 5G is key to creating more cost-effective telcos explains more.

We will go deeper into aspects of the technology involved and the research and standardization being done to continuously improve network energy performance, minimizing network energy consumption for a particular performance and capacity.

NR has improved the potential for network energy savings, but excessive energy is still spent at times when a network serves no or little traffic, and consequently, the energy is wasted. We are engaged in 3GPP to enable even lower energy consumption of NR at both low-, medium and high-load scenarios. The variation in User Equipment (UE) performance due to traffic load is often larger than the variation due to activation of the studied energy saving techniques. Hence, for the former, the variations in UE performance can be significantly reduced by applying energy saving features during low and medium traffic. For the latter, although energy consumed for service provisioning in high traffic load scenarios may be seen as justifiable, energy saving techniques in spatial-, time-, power-, and frequency domains can still reduce the energy consumed by NR base stations – gNB (Next Generation Node B) also in high traffic load scenarios.

Network energy consumption modelling

As a foundation for this work, we recently proposed a network energy consumption model which captures the impact of the gNB switching on/off selected components, for example, antenna chains and digital units. Our model has helped significantly in shaping one of the agreed-upon network energy consumption models in 3GPP. This blog post is based on contribution papers submitted to 3GPP by Ericsson, proposed in RAN1#111.

In our energy consumption model, sleep mode is defined as when one or many components are switched off for energy saving. Active mode is when all components are switched on. Transition time or transition duration is a concept we introduce for time needed to ramp up or down from a sleep state to an active state. Depending on which components are turned off and the associated mode transition duration, sleep mode is divided into three states: microsleep, light sleep, and deep sleep.

The definition of each sleep state is given below:

Microsleep

In this state, the network turns off components with fast transition times, so that reactivation may occur essentially instantaneously, sometimes as fast as a few microseconds. Examples of such components are power amplifiers (PAs) and low noise amplifier (LNAs).

Light sleep

This is a state in which additional circuitry components with medium transition time (several milliseconds) are turned off, for example, integrated circuits (ICs) in transmitter and receiver units.

Deep sleep

In this state, the network turns off as many circuitry components as possible and keeps only a few components on, for example, internal clocks, with long transition times at up to several tens of milliseconds.

The model also defines active modes for uplink (UL) and downlink (DL), where all available hardware resources are utilized. The transition time of the different sleep states – microsleep, light sleep, and deep sleep – are shown in Table 1. Clearly, the more components that are turned off and the deeper they sleep, the longer it takes to wake up the sleeping gNBs.

The operating power levels of these states are also given in Table 1. Since different network vendors may adopt different implementations on gNBs, the numbers in the model are relative rather than absolute values, and the reference is the power levels at the deep sleep state.

The model also considers scaling based on transmission and reception bandwidth, carrier aggregation (CA), and the number of antenna receiver units (RUs). Table 2 shows how the power of gNB is scaled depending on the scaling factor.

Here we can see that changing different factors will result in significant energy saving on a gNB. For example, reducing or scaling down the number of antennas from 64 to 32 can enable a 30 percent energy saving, as shown in the last column of table 2. Therefore, this energy model leads us to create different network energy saving techniques which will be discussed in the next section.

Our model considers optimized hardware and software components and ongoing technology advances, ensuring that it will apply as a basis for energy performance evaluation of the network in the future, for example, within a five to six-year time horizon, even after the software or hardware has undergone considerable evolution.

Network energy savings techniques

In addition to putting a gNB into sleep mode when there is no activity, other energy saving techniques are also required for gNBs in the active state. In the following figures, we illustrate several such example technologies in spatial, power, and time domains and evaluate them to quantify their potential effect on network energy saving.

Dynamic antenna adaption (spatial domain)

gNBs with active antenna systems generally employ a large number of antenna elements arranged in subarrays as exemplified in Figure 1. Each subarray is then typically connected to two transceiver chains, where each transceiver chain is further connected to one of the polarizations.

Figure 1: Exemplary antenna arrangement (A) of a gNB with an active antenna system including subarrays of four antenna elements with two polarizations. A typical setup (B) where each polarization of the subarray is connected to a separate receiving (Rx) and transmitting (Tx) chain. Source: R1-2212155, Network energy savings techniques, Ericsson, 3GPP TSG RAN WG1 #111

Commonly, gNBs are equipped with arrays of 64 or more Receiving (Rx) and Transmitting (Tx) chains, and even larger arrays are foreseen in the future, especially at higher frequencies. The energy consumed by this multitude of transceiver chains corresponds to a major part of the total consumed gNB energy. Efficient beam management also requires a higher number of reference signal transmissions such as Channel State Information Reference Signals (CSI-RSs) which, due to excessive signaling and reduced sleep opportunities, contribute to high energy consumption.

Not all the large array transmissions and the activated transceiver chains are needed all the time. Savings can be achieved by muting a portion of the transceivers and turning off the involved circuitry as shown in Figure 2.

Figure 2: Different transceiver muting patterns depending on the deployment, load, and coverage of UEs. Source: R1-2212155, Network energy savings techniques, Ericsson, 3GPP TSG RAN WG1 #111

The number of transceivers to be muted depends on the deployment, load, and UE coverage scenarios, which makes it necessary for the gNB to dynamically adapt the muting pattern. There can also be a trade-off between gNB energy saving gains and user performance loss, e.g. user throughput. To avoid such loss due to recurrent reconfigurations and transceiver muting, the gNB needs to acquire knowledge of the potential user performance impact of the different muting patterns prior to the actual transceiver muting decision. This requires the UE to report not only Channel Status Information (CSI) for the current transceiver configuration, for example, 64 chains, but also CSI for other candidate configuration(s), for example, 48, 32, or 16. In Figure 3, we evaluate the effect of different antenna configurations on energy consumption by varying the number of transceiver chains of a gNB between 16, 32, and 64. Energy consumption and user throughput are investigated at three typical system loads (occupied physical resources) where low-, light-, and medium loads represent up to 15, 30, and 50 percent system resources, respectively.

As the figure illustrates, fewer antenna chains lead to lower energy consumption under any load. When the network load is low, the network performance loss caused by reducing the antenna chains from 64 to 16 is below three percent for the majority of the UEs. Even at medium load, the performance loss of 16 chains is only about seven percent, while the energy consumption is reduced to about 61 percent compared to 64 chains, which is a highly significant benefit.

We used a semi-dynamic antenna configuration, and it can be predicted that if we dynamically configure the network according to certain indicators or thresholds, we can minimize the loss of network performance while obtaining excellent energy-saving performance.

Dynamic power adaption (power domain)

Currently, the DL transmission power is fixed by the gNB for all UEs in a cell in a way such that any UE at the cell edge doesn`t suffer from too low signal-to-interference-plus-noise ratio (SINR) levels which would seriously impact its ability to receive network signaling. In contrast, UEs in the cell center may experience over-dimensioned SINR that does not improve perceived performance. In such scenarios, gNB energy performance is improved by reducing the gNB transmission power specifically for the UEs in good coverage regions where a lower SINR would not affect the UE’s perceived performance as shown in Figure 4.

Figure 4: Exemplary power arrangement of a gNB (A) with UEs in the cell center and at the cell edge.

It is intuitive that, in the absence of any active UEs at the cell edge, user throughput might not be significantly affected when reducing the transmit power of the gNB to a level ensuring that the SINR at UEs' locations still stays favorable, while the energy consumption of the gNB is reduced. To verify this, we evaluated the effect of different power configurations on energy saving by varying the transmission power of a gNB between 43 dBm, 49 dBm, and 55 dBm in Figures 5 and 6. Energy consumption and average user throughput are again investigated at three typical system loads.Figure 5 illustrates that when no UEs are deployed near the cell edge, lower transmission power leads to reduced gNB energy consumption in all load conditions.

Moreover, the higher the system load, the more evident the energy-saving effect of reducing the transmission power. Meanwhile, average user throughput is hardly affected by different transmission power configurations. Clearly, the UE distribution in this scenario is favorable for energy saving by transmission power reduction.

On the other hand, Figure 6 depicts a scenario where UEs are located near the cell edge and indicates that, in such a scenario, reducing gNB transmission power seriously affects average user throughput. Therefore, this technique needs to be selectively applied in appropriate cells.

Dynamic configuration adaptation (time domain)

According to the energy model described above, putting a gNB into sleep mode when it is in idle mode is an effective energy-saving method. Under the premise of not affecting network performance, the longer the continuous idle mode, the more opportunities the gNB has to enter and maintain a deeper sleep state, thereby saving energy consumed by the gNB. Therefore, creating more chances of idle mode within the time domain is a direct way of improving energy performance. To achieve this goal, we can dynamically adjust certain configuration to provide more sleep opportunities for the gNB during idle mode while ensuring network performance during non-idle mode.
For instance, NR is designed with a typical synchronization signal block (SSB) periodicity of 20 ms-160 ms. At first glance, it seems that if there is nothing to transmit or receive between two SSB transmissions, the gNB should be able to go to sleep and save a significant amount of energy. However, this may not be possible due to periodic reception occasions for the gNB. For example, the gNB needs to at least listen to Physical Random Access Channel (PRACH) preambles for potential UEs accessing the cell even if no explicit UL transmission grants have been given. However, the reception period of PRACH has multiple options and can also be reconfigured. Therefore, if we dynamically configure the reception period of PRACH based on the network status, we can provide more sleep opportunities for the gNB in idle mode.

Figure 7 visualizes the total energy consumption and its distribution between different operations, sleep states, and state transitions for a gNB in idle mode performing SSB and Remaining Minimum System Information (RMSI) transmissions every 20 ms.

Figure 7: Energy consumption comparison between four setups, with 10 ms, 20 ms, 40 ms, and 80 ms PRACH occasion periodicity (left to right). Plotting tools: MATLAB R2023a, Simulation Engineer: Ericsson Internal Simulator

The four configurations differ in terms of gNB receiver activity. The baseline configuration is optimized for performance where the PRACH occasion period is set to every 10 ms so that the initial connection establishment delay is minimized. In alternative configurations, the PRACH occasion periodicity is increased to 20/40/80 ms respectively to improve gNB energy performance. As can be seen in the figure, reducing PRACH preamble monitoring in the gNB brings considerable energy savings. Extending the PRACH occasion period from 10 ms to 40 ms can save up to 21 percent in energy consumption. However, extending the PRACH occasion period beyond 80 ms becomes less attractive considering the limited additional energy saving gain versus the higher access delay of UE, especially at higher access loads. Therefore, it is beneficial for both network and UE energy savings if the PRACH occasion period can be dynamically configured depending on the use cases, for example, some idle UEs need a faster transition to connect, preferably by more light-weight mechanisms than a System Information update which itself incurs a delay that cannot be ignored.

Breaking the energy curve

With energy prices soaring and the demand for mobile networks continuing to grow, energy use and related emissions will also increase unless we act. Our network energy consumption model can predict the energy consumption for both current and future networks, and additionally enhance the current NR mechanisms to provide more energy savings. By reducing energy use and related emissions, we further contribute to being able to break the energy curve in mobile networks.

Further reading

Download our 3GPP RAN1#111 submission paper in full here: “Evaluations for network energy savings techniques”, and “Network energy savings techniques”.

Read the blog post Crush the energy curve! How 5G is key to creating more cost-effective telcos

Read this report on 5G energy performance that highlights how 5G can become more energy-efficient.

Read more about 5G New Radio (NR)

Read more on Energy Sharing: Energy from everywhere