RESUMO
With the increasing utilization of data in various industries and applications, constructing an efficient data pipeline has become crucial. In this study, we propose a machine learning operations-centric data pipeline specifically designed for an energy consumption management system. This pipeline seamlessly integrates the machine learning model with real-time data management and prediction capabilities. The overall architecture of our proposed pipeline comprises several key components, including Kafka, InfluxDB, Telegraf, Zookeeper, and Grafana. To enable accurate energy consumption predictions, we adopt two time-series prediction models, long short-term memory (LSTM), and seasonal autoregressive integrated moving average (SARIMA). Our analysis reveals a clear trade-off between speed and accuracy, where SARIMA exhibits faster model learning time while LSTM outperforms SARIMA in prediction accuracy. To validate the effectiveness of our pipeline, we measure the overall processing time by optimizing the configuration of Telegraf, which directly impacts the load in the pipeline. The results are promising, as our pipeline achieves an average end-to-end processing time of only 0.39 s for handling 10,000 data records and an impressive 1.26 s when scaling up to 100,000 records. This indicates 30.69-90.88 times faster processing compared to the existing Python-based approach. Additionally, when the number of records increases by ten times, the increased overhead is reduced by 3.07 times. This verifies that the proposed pipeline exhibits an efficient and scalable structure suitable for real-time environments.
