Abstract

This paper describes SUSTAIN-CTRL, an information systems (IS) sustainability artificial intelligence (AI) control framework that places an LSTM-based forecasting model at its "heart". The framework processes heterogeneous telemetry data, including computation, environmental, and network sensors, as well as unstructured logs, transforming them into synchronized feature vectors. Fine-grained, multi-stream, energy-demand forecasting, semantic inefficiency scoring, and anomaly detection at a parallel inference stage using LSTM, transformer, and autoencoder models, respectively, generate labels for a consolidated decision dataset. This dataset is used by a heuristic multi-objective optimizer that minimizes a weighted sum of energy consumption, SLA-violation penalties, and carbon emissions. The carbon budget and SLA-validated control vectors are optimized while generating human-readable explanations for auditability. SUSTAIN-CTRL issues dynamic VM scaling and cooling command execution via orchestration APIs, closing the real-time feedback loop for continuous adaptation. Empirical evaluation using a 24-hour testbed showed that SUSTAIN-CTRL reduced energy consumption and emissions by 22.7% and 14.5%, respectively, while improving SLA by 2.7 percentage points. These results confirm that LSTM forecasting, when integrated into a closed-loop AI architecture, sustains performance-aligned operation decoupled from LSTM forecasting.