QBSD: Quartile-Based Seasonality Decomposition for Cost-Effective RAN KPI Forecasting

TL;DR

This article introduces QBSD, a live single-step forecasting approach tailored to optimize the trade-off between accuracy and computational complexity that excels in runtime efficiency compared to the leading algorithms available while maintaining competitive forecast accuracy that rivals neural forecasting methods.

Abstract

Forecasting time series patterns, such as cell key performance indicators (KPIs) of radio access networks (RAN), plays a vital role in enhancing service quality and operational efficiency. State-of-the-art forecasting approaches prioritize accuracy at the expense of computational performance, rendering them less suitable for data-intensive applications encompassing systems with a multitude of time series variables. They also do not capture the effect of dynamic operating ranges that vary with time. To address this issue, we introduce QBSD, a live single-step forecasting approach tailored to optimize the trade-off between accuracy and computational complexity. The method has shown significant success with our real network RAN KPI datasets of over several thousand cells. In this article, we showcase the performance of QBSD in comparison to other forecasting approaches on a dataset we have made publicly available. The results demonstrate that the proposed method excels in runtime efficiency compared to the leading algorithms available while maintaining competitive forecast accuracy that rivals neural forecasting methods.

Figures (5)

• Figure 1: Schematic architecture of QBSD forecast computation, where $t$ is the current timestamp and $k$ is the context period.
• Figure 2: Flowchart illustrating QBSD. The flow begins with the input timestamp for which the forecast should be generated. This forecast along with the observed KPI value of the given timestamp is used for residual computation.
• Figure 3: Sample input data. Percept history is highlighted in green while QBSD is performed on the example sequence highlighted in blue. Note that there can be missing data in history.
• Figure 4: Illustration of QBSD applied on the example highlighted in Fig. \ref{['fig:sampleinput']} along with normalized residuals. Red bars indicate definitive anomalies wherein the blue bar indicates a peak that is not statistically an anomaly.
• Figure 5: Plot of actual KPI values (A through F) along with forecast, $Q_1$, and $Q_3$ obtained using the QBSD algorithm for the Cell-F dataset. Series $Q_1$ and $Q_3$ have been smoothed with the Savitzky–Golay filter.