Study on cloud data center resource consumption prediction based on BILSTM, ARIMA, and SVR

With the rapid advancement of cloud computing, accurately predicting resource consumption in cloud data centers has become a critical task, crucial for enhancing resource management and optimizing energy efficiency. This paper proposes a novel hybrid framework, Bidirectional Long Short-Term Memory–Autoregressive Integrated Moving Average–Support Vector Regression (BILSTM–ARIMA–SVR, hereafter BAS), which integrates the strengths of BILSTM, ARIMA, and SVR. In the BAS approach, the three sub-models are trained individually, and a genetic algorithm is employed to optimize their weighted fusion. Experimental results demonstrate that BAS achieves superior overall performance compared to other benchmark models — including CNN–LSTM, BILSTM, SVR, and ARIMA. While CNN–LSTM attains a slightly lower MAE, its MSE is unusually high and its stability is poor, whereas BAS consistently delivers lower MSE, higher [Formula: see text], and greater robustness, particularly when dealing with complex and volatile time-series data. These findings confirm that BAS enhances both prediction accuracy and robustness, thereby facilitating effective resource scheduling and energy-saving optimization. Future research will focus on improving computational efficiency and exploring the feasibility of applying BAS to large-scale, real-time prediction scenarios.