МАШИНСКО УЧЕЊЕ ЗА ПРЕДИКЦИЈУ ЕНЕРГЕТСКЕ ПОТРОШЊЕ У IOT ОКРУЖЕЊУ: АНАЛИЗА LSTM И TCN МОДЕЛА

Real-time data processing has become the foundation for a wide range of IoT applications running across edge and cloud resources within the computing continuum. However, such applications typically generate high operational costs due to the use of cloud-managed messaging services (e.g., AWS SQS, GCP Pub/Sub) for data collection, or due to the deployment of AI-enabled IoT devices that support federated learning at the edge.

To reduce these costs, this paper proposes an Adaptive Predictive System (APS). APS trains machine learning models in the cloud to perform real-time data prediction for IoT applications. For model training purposes, APS relies on lightweight processing at the IoT device, which performs downsampling before transmitting data to the cloud. During operation, APS measures prediction error and dynamically adjusts the sampling rate to ensure minimal prediction deviations in accordance with application requirements. In addition, APS applies inference-time correction to prevent gradual model degradation and drift.

This approach eliminates the need for expensive AI-enabled IoT devices, reduces the volume of data transmitted to the cloud via messaging services, and maintains low prediction error. To evaluate the effectiveness of the system, we developed an APS prototype and tested it on real-world IoT data. The results demonstrate promising outcomes, including an approximate 80% cost reduction with an average prediction error of around 5%.