Attention-based integrated deep neural network architecture for predicting the effectiveness of data center power usage

Addressing the critical need for enhanced power usage effectiveness in data centers (DCs), this study pioneers an improved convolutional long short-term memory with deep neural network (CLDNN) model, enriched with attention mechanisms for precise DC performance prediction. We rigorously evaluate our model against leading architectures – long short-term memory (LSTM), attention-based (att-LSTM), convolutional LSTM (CNN-LSTM), gated recurrent unit (GRU), and CNN-GRU – to affirm its superiority in predictive accuracy and robustness. The integration of convolutional layers processes hourly data inputs efficiently, reducing complexity and improving pattern detection. A subsequent flattening layer optimizes accuracy, while a dual-layered LSTM and a deep neural network delve into frequency, temporal dynamics, and complex data relationships. Incorporating an attention mechanism into the att-CLDNN model has revolutionized predictive analytics in DC energy management, significantly enhancing accuracy by highlighting crucial data interdependencies. This model's unparalleled precision, evidenced by achieving the lowest Mean Squared Error (MSE) of 0.000179, the minimum Mean Absolute Error (MAE) of 0.01048, and the highest R2 Score of 0.977031, underscores its effectiveness. Crucially, this breakthrough fosters sustainability in energy management, promoting greener DC operations through precise energy use predictions, leading to substantial energy savings and reduced carbon emissions, in alignment with global sustainability objectives.