Artificial Intelligence Workloads Forecasting based on Correlation Multi-channels Fusion

Time series forecasting methods are largely used to predict the resource demands of AI workloads in Cloud Computing environments, thereby achieving optimized resource allocation and effective energy conservation. Despite existing efforts of time-series prediction, the accuracy of time series forecasting is often limited by the complexity and dynamics of the input features, especially in multi-channel AI workloads, where the frequent interaction and complex correlation between input features significantly affect the forecasting results. Consequently, a detailed analysis of the correlation between multi-channel input features of AI workloads and their effect on forecasting performance is vital for enhancing the accuracy and reliability of time-series forecasting. We conduct an in-depth analysis on the intricate correlations among multiple input features of AI workloads based on multi-channel inputs using the Alibaba Cloud trace logs to investigate the impacts of such correlations on prediction accuracy. Moreover, we design a novel module, namely a Feature Selection Module (FSM), to fuse highly correlated feature channels, thereby extracting intrinsic correlation information to further improve prediction accuracy. In a 21-feature prediction experiment, our enhanced design, integrating FSM with a Long Short-Term Memory (LSTM) time-series model, achieves 85.71% better results in terms of MAE,MSE and R2 than existing models including LSTM and transformer.