Intelligent O&M Based on AI Power Saving

High electricity charges have always been a big issue for operators. According to earlier data, the annual electricity expense of China's three major operators (China Mobile, China Unicom, China Telecom) even exceeded 30% of their employee’ salaries. In the coming 5G era, more wireless base stations will be deployed. Test results from China Mobile show that the power consumption of 5G base stations is several times higher than that of 4G base stations, which may bring huge electricity costs. In the beginning of 2019, China Telecom made four measures to reduce costs, the first of which was to reduce electricity fees.

The number of base stations accounts for 70% of devices in a mobile network. With the increase in coverage, capacity, and the number of sites, the power consumption of base stations accounts for more than half of the mobile network, even up to 80%. Therefore, site energy saving is the top priority of power saving for operators. Reasonable network planning and NE-level power saving are two important aspects of ZTE's non-AI power saving solutions:

—Reasonable network planning means that through network planning and optimization, the minimum number of sites are deployed to meet the requirements of network capacity and power saving. Coverage efficiency is improved while invalid system overheads can also be reduced.
—NE-level power saving can be applied to more scenarios. Multi-layer shutdown including cell shutdown, channel shutdown, and symbol shutdown, which is widely used all over the world, is the most effective power saving solution for traditional network elements.

By June 2019, NE-level power saving solution had been commercially used in nearly 40,000 sites with China Telecom and China Unicom, and almost 20,000 sites overseas had also enabled the solution. According to statistics, up to 10% power consumption are saved, bringing a considerable Opex saving to operators.

However, challenges of the non-AI power saving solution are also obvious. First, the solution is not flexible. To simplify the deployment complexity, unified parameters would be configured in one region even the whole network without scenario identification, which may be poor matched to the real traffic variation. Second, the solution does not support sustainable evolution. Once there are changes in network architecture, such as the increase or decrease in the number of cells, the changes in neighbor cell relationship, or the changes in traffic, the previous power saving policy may be completely invalid. Third, the solution has poor adaptability. A large number of field or remote support personnel are needed. They speed much more time in analyzing KPIs, traffic load, and trial results.

It is therefore necessary to develop an intelligent and self-adaptive power saving solution. With the help of three AI capabilities such as data perception, AI analysis, and intent insight, AI accelerators are introduced to provide rapid AI training and reasoning for intelligent O&M where power saving and network optimization are particularly important.

Through network prediction, strategy adjustment and optimization, and real-time KPI monitoring, the AI power saving solution forms a closed loop that can find a balance between power saving and network performance with little labor cost.

Time-based power saving:

Network load and user behavior can be predicted through historical data analysis. Autoregressive integrated moving average model (ARIMA), long short-term memory (LSTM) and Facebook's Prophet are used for time series prediction. Instead of setting up unified parameters, cell-based time window ensures more activation time for power saving.

Threshold-based power saving: Different triggered thresholds are used for different scenario through prediction of RRC connected users and PRB usage. Comparing the predicted value with the real-time traffic based on 15 minutes granularity, the prediction accuracy exceeds 90%, which means strategy and scenario recognition cell by cell improves enforceability.
KPI-based power saving: Network KPIs such as call setup success rate, drop rate, handover success rate and throughput, average power consumption and user experience are monitored in real time for policy optimization. Real-time policy roll-back is allowed in AI power saving. Gains of power saving improve with little impact on the network.

ZTE and China Unicom has deployed the AI power saving solution in Shandong, China, since June 2019. The commercial trial involves three phases. In the Phase I,167 cells including LTE single mode, GL multi-mode, and dual-band LTE enabled AI power saving. In the Phase II, more than 1,000 cells enabled AI power saving, both time window and threshold adjustment were trialed. In the Phase III, more than 10,000 cells will enable AI power saving in batches. Cell shutdown, channel shutdown, and symbol shutdown are all trialed based on site configuration.

The two trials verify that time for power saving activation has increased by 50% to 80%, and more than 10% to 30% power has been saved compared to the non-AI solution. Almost 3 kWh/day per site has been saved in Shandong, China.

Based on the trial results, we can make a simple calculation: With the AI power saving, 2 to 3 kWh/day per site would be saved in Shandong, China. When 10,000 cells (3300 sites) enables the AI power saving, 2.5 million kWh of electricity will be saved each year. If the average cost of commercial electricity is 1 yuan/kWh, 2.5 million yuan can be save to China Unicom in Shandong only. To operators, more real money is saved based on the AI power saving.

Of course, network-wide intelligence is difficult to achieve overnight, and it needs to go through a long-term development. However, with the continuous accumulation of commercial data samples, the best self-learning algorithm for AI power saving could be gradually approached. AI power saving will also be evolved to obtain more sample points and improve AI algorithms. It will help to reduce the network carbon footprint and create a greener intelligent network in the future.