BUSINESS | NER

Next-Generation Forecasting Engine, PatchTST: To overcome the long-range data limitations of conventional RNN- and LSTM-based forecasting models,
we adopted PatchTST (Patch Time Series Transformer), a state-of-the-art Transformer architecture
Multivariate Time-Series Patch Learning: Data Patching: Structures multivariate data such as PM and CO2 into time-based patches

Feature Analysis: Learns local correlations and long-term trends simultaneously to maximize analytical efficiency

Forecasting Performance: Predicts pollutant concentration trends over the next 3 hours with over 90% accuracy
Low-Power, High-Speed Edge AI Engine Optimization: Model Optimization: Lightweight PatchTST through quantization and pruning

MCU Deployment: Improved efficiency for stable computation even on low-spec devices

Ultra-Low-Latency Inference: Independent edge inference within 100 ms, with no cloud round-trip

Low-Power Operation: Minimized inference power consumption for maximum energy efficiency