Energy-Efficient Optimized Dynamic Massive MIMO Based on Predicted User Quantity by LSTM Algorithm

Abstract

Massive multiple-input multiple-output (MIMO) is one of enabling promising techniques to meet 1000X capacity increase of the future communication network with high efficiency. As the network load varies dynamically throughout the day, the base stations (BSs) often waste a lot of energy when the traffic loads are low. The smarter resource allocation scheme always brings about higher energy efficiency (EE). Obtaining the future variation of network load at different time intervals is the key to resource allocation. In this paper, an EE optimized scheme over predicted traffic loads is proposed to deal with the energy inefficiency problem resulted from the high fluctuations of traffic loads. We predict the variation of the number of users associated to a BS equipped with massive MIMO antenna array and allocate appropriate number of antennas based on the prediction result to maximize the downlink EE. We use the long short-term memory (LSTM) algorithm to perform the predication based on a real dataset from a major telecom operator in China. Our LSTM scheme is also compared with other machine learning algorithms and it achieves higher prediction precision. Employing our prediction method and dynamically adjusting the number of antennas, we can get 101% gain in EE compared to a baseline system where the BS runs with the fixed number of antennas all time.