Abstract
As global industries strive toward "double carbon" goals—carbon peaking and carbon neutrality—cloud data centers have emerged as critical targets for energy optimization. These facilities often suffer from a mismatch between resource supply and demand, leading to significant energy waste and carbon emissions. This paper presents a hybrid deep learning model, CNN-BiLSTM, designed to predict cloud computing loads with high precision. By integrating spatial feature extraction via Convolutional Neural Networks (CNN) and temporal dependency modeling via Bidirectional Long Short-Term Memory (BiLSTM) networks, the proposed method enables dynamic carbon emission forecasting. Experimental results using Google Cluster Data demonstrate that the CNN-BiLSTM model reduces Mean Square Error (MSE) by up to 52% compared to traditional models, providing a robust framework for sustainable data center management.
The Challenge of Data Center Carbon Emissions
Data centers are massive energy consumers, with power consumption distributed across IT equipment, cooling systems, and power distribution units. A primary inefficiency in these environments is the static allocation of resources; servers often remain idle or underutilized, yet continue to consume power and require cooling.
As shown in Figure 2, IT equipment and cooling systems account for approximately 85% of total power usage. To achieve meaningful carbon reduction, data center managers require tools that can predict load fluctuations in real-time, allowing for the dynamic adjustment of power and cooling systems based on actual demand rather than static overheads.
Methodology: Capturing Spatial-Temporal Duality
The proposed CNN-BiLSTM architecture addresses the inherent complexity of cloud workloads, which exhibit both spatial and temporal characteristics.
- Spatial Feature Extraction (CNN): The model utilizes a CNN to identify correlations between different Machine IDs and Task IDs. This "spatial" dimension captures how different nodes in a cluster interact and share the workload.
- Temporal Modeling (BiLSTM): Unlike standard LSTMs, the BiLSTM processes data in both forward and backward directions, capturing long-term dependencies and sudden "bursty" patterns typical of cloud applications.
- Dynamic Carbon Modeling: Predicted CPU utilization is converted into real-time power consumption using server-specific formulas (e.g., Intel specifications). This power value is then translated into carbon emissions using the national power grid's emission factors (e.g., 0.5703 t CO2 / 10,000 kWh).
Figure 4 illustrates the dual nature of the data: the temporal fluctuations over time (a) and the spatial distribution across different machine IDs (b). The hybrid model is specifically designed to exploit these multi-dimensional patterns.
Experimental Results and Discussion
The model was validated using the Google Cluster Data set, a large-scale trace of approximately 12,000 machines. The CNN-BiLSTM model significantly outperformed several benchmarks in terms of prediction accuracy:
- MSE Reduction: Compared to the BP neural network, standard LSTM, BiLSTM, and CEEMDAN-ConvLSTM models, the MSE decreased by 52%, 50%, 34%, and 45%, respectively.
- Fitting Accuracy: The model achieved an R-squared (R2) value of approximately 0.95, indicating a near-perfect correlation between predicted and actual load values.
- Carbon Impact: In a 24-hour simulation, the dynamic prediction method calculated a footprint of 3.12 kg of CO2, compared to 3.81 kg estimated by traditional static methods. This suggests that dynamic management can prevent the overestimation of energy needs and facilitate more efficient cooling cycles.
Conclusion
The CNN-BiLSTM model provides a highly accurate mechanism for predicting cloud computing loads and their associated carbon footprints. By moving from static power models to dynamic, spatial-temporal predictions, data centers can significantly reduce their environmental impact while maintaining quality of service. Future research will focus on integrating attention mechanisms to further refine prediction during extreme workload spikes and exploring federated learning to address data privacy concerns in multi-tenant cloud environments.
Summary updated